Académique Documents
Professionnel Documents
Culture Documents
Artificial societies
Artificial societies
Downloaded by [Universidad Autnoma de Madrid] at 04:28 05 November 2013
Nigel Gilbert
University of Surrey
&
Rosaria Conte
Institute of Psychology
Italian National Research Council
Contents
Preface
vii
Notes on contributors
ix
122
132
160
180
240
Index
252
Preface
This volume draws on contributions made to the second of a continuing series of
international symposia on simulating societies. All the chapters, except the one by
Robert Axelrod, first saw the light of day at a meeting held in a converted monastery at
Certosa di Pontignano, near Siena in Italy, in July 1993. Robert Axelrod was not able to
be present, but contributed a chapter nevertheless. The meeting in Siena followed one
held at the University of Surrey, Guildford, UK, in April 1992 (see Gilbert and Dorans
Simulating societies: the computer simulation of social phenomena, 1994). The editors
and contributors are very grateful to Cristiano Castelfranchi and Rosaria Conte for their
hospitality in arranging the Siena meeting.
As we note in the Introduction (Chapter 1), this volume and its predecessor provide
evidence for the growth and current liveliness of the field. The vast power of current
personal computers and the sophistication of the software and the programming tools
now available means that increasingly the limit on the computer simulation of societies is
not the difficulty of implementing simulations, but the imagination and creativity of the
researcher. That this imagination is not lacking is shown by the range of contributions to
this volume, which draws on almost the whole span of the social sciences to learn more
about and experiment with everything from tastes for cultural objects to patterns of
Aboriginal kinship, and from the organization of ant nests to the location of cities.
The editors are grateful to Justin Vaughan and Kate Williams of UCL Press for their
encouragement to embark on this series and for their patience in helping to put together
the book, and to Jim Doran for his help in organizing the Siena conference. We also
gratefully acknowledge the support of the Universities of Siena and Surrey.
NIGEL GILBERT ROSARIA CONTE July 1994
Notes on contributors
Robert Axelrod is the Arthur W.Bromage Distinguished University Professor of
Political Science and Public Policy at the University of Michigan. He received his BA
in mathematics from the University of Chicago, and his PhD in political science from
Yale University. He is the author of The evolution of cooperation (1984) and Conflict
of interest (1970), the editor of Structure of decision (1976), and the author of
numerous articles including An evolutionary approach to norms (1986). His work
on cooperation and norms has received awards from the American Association for the
Advancement of Science, the American Political Science Association, the MacArthur
Foundation and the National Academy of Sciences. He has been elected to
membership in the American Academy of Arts and Sciences, and the National
Academy of Sciences.
Stphane Bura is writing a doctoral thesis in distributed artifical intelligence at the Paris
VI University. His research is focused on the emergence of structures in complex
systems. His interests include artificial life, social simulation and memetics.
Cristiano Castelfranchi is a researcher at the Institute of Psychology of the National
Research Council of Italy. He has been the head of the Project for the Simulation of
Social Behaviour since 1988. He is the author of several books on social and cognitive
modelling. In 1992 and 1993 he was Programme Chairman of the European Workshop
on Modelling Autonomous Agents in a Multi-Agent World (MAAMAW).
Federico Cecconi is with the Institute of Psychology of the National Research Council in
Rome. He has a technical degree in computer science. He is interested in neural
networks and genetic algorithms and in their application for both scientific and
technological purposes.
Antonio Cerini did his thesis work for a laurea in philosophy at the Institute of
Psychology of the National Research Council in Rome. He is currently with ISED
Informatica, a software house in Rome.
Rosaria Conte is Researcher at the Institute of Psychology of the Italian National
Research Council (CNR), Rome, and is teaching Cognitive and Social Psychology at
the University of Torino. Her first degree is in philosophy and her post-graduate
research training is in the cognitive modelling of social behaviour. She coordinates a
research group on the simulation of social behaviour at the Institute of Psychology,
with the aim of working out computational models of social phenomena both at the
micro and at the macro levels. Her interests include cognitive modelling, social
simulation, multi-agent systems, distributed artificial intelligence and the theory of
rational action.
Bruno Corbara has a doctorate in Ethology from the University of Paris XIII. His first
degree is in psychology and his doctorate is on the social organization and its genesis
in the Ectatomma ruidum ant. He has studied many species of social animals and has
published two books on social insects. He is currently writing a book on the various
building behaviours, either collective or individual, that exist in the animal world. His
interests include self-organized systems, theories of behaviour, and statistical methods
for ethological research.
Jim Doran read mathematics at Oxford, became a specialist in artificial intelligence at
the University of Edinburgh and is now Professor at the University of Essex.
Throughout his career he has applied statistical and computer methods in archaeology.
His particular interest is in the simulation of societies using the concepts and
techniques of distributed artificial intelligence.
Alexis Drogoul is Assistant Professor at the University of Paris VI. His first degree is in
computer science from the University of Paris VI and his doctorate is on the use of
multi-agents systems in ethological simulation and problem solving. The main theme
of his work concerns the ability of groups of agents to self-organize in order to
produce functional collective responses, and more particularly, the emergence of these
responses within groups of non-intelligent agents (also called reactive agents). His
interests include biological simulation, theories of the emergence of cognition, and
distributed problem solving.
Massimo Egidi is Professor of Economics at the Department of Economics of the
University of Trento, Italy. He is especially interested in the theory of organization
and collective decision making.
Nicholas V.Findler is Research Professor of Computer Science and the Director of the
Artificial Intelligence Laboratory at Arizona State University, Tempe, Arizona, USA.
He has a BE (Honors, Summa-cum-Laude) in electrical engineering and a PhD in
mathematical physics from the Technical University of Budapest, Hungary. He has
worked in artificial intelligence since 1957. A large proportion of his projects over the
past 12 years or so have been in the area of distributed artificial intelligence, ranging
from applications in future air traffic control, street traffic signal control, through
allocation of moving resources to moving tasks and nationwide manufacturing control,
to studies of the intelligent and heterogeneous multi-agent systems described in this
book.
Michael Fisher is a Principal Lecturer in Computing Research at the Manchester
Metropolitan University. His first degree was in mathematics and computer science,
and for his PhD he worked on the temporal semantics of concurrent programming
languages. He is well-known for his work in both the theory and implementation of
temporal theorem-proving, and is one of the leading proponents of executable
temporal logics as programming languages. His other research interests include DAI,
formal specification and verification, object-based systems, and adult learning.
Nigel Gilbert is Professor of Sociology at the University of Surrey. His first degree is in
engineering from the University of Gambridge and his doctorate is on the sociology of
scientific knowledge. He is a founder member of the Social and Computer Sciences
research group at Surrey, which applies insights from the social sciences to the design
of information technology. His interests include social simulation, natural language
interaction with computer systems, and statistical methods for social research.
France Gurin-Pace is a demographer and a statistician at the French National Institute
for Demographic Studies (INED). She has written Deux sicles de croissance urbaine
(Two centuries of urban growth) Anthropos Ed.
Psychology. His research interests include neural network and artificial life models to
study the behaviour of simple and, hopefully, more complex organisms.
Denise Pumain is a geographer and a professor at the Paris I University. She is a
technical consultant at the French National Institute for Demographic Studies (INED)
and director at the CNRS (National Center for Scientific Research). She has written
books about urbanization and modelling. She is the publishing director of the Villes
collection (Anthropos Ed.).
Lena Sanders is a demographer and a statistician in the PARIS (Pour lAvancement de
la Recherche sur lInteraction Spaciale) team at CNRS. She teaches geography at the
Paris VII University. She has written a book on data analysis in geography and cowrote Villes et auto-organisation with D. Pumain and T.Saint-Julien.
Jean Pierre Treuil is in charge of the Laboratory of Computer Sciences at the French
research institute, ORSTOM. He is a graduate of the Paris Ecole Polytechnique and
Ecole Nationale de Statistique et dAdministration Economique. His doctorate is in
artificial intelligence and learning processes. His interests include artificial life,
methods of distributed artificial intelligence and computer simulations of social
processes.
Michael Wooldridge is a Lecturer in Computing at the Manchester Metropolitan
University. He obtained a first-class honours degree in computer science in 1989, and
went on to study for a PhD, which he obtained from UMIST in Manchester in 1992.
His doctoral research was in the area of logics for modelling distributed AI systems.
His current research interests lie primarily in the area of distributed AI, and he has
publications in both the theoretical and practical aspects of the field.
Chapter 1
Introduction
Computer simulation for social theory
Rosaria Conte and Nigel Gilbert
In the Introduction to this volumes predecessor, Simulating societies: the computer
simulation of social phenomena (Gilbert & Doran 1994), the editors, Nigel Gilbert and
Jim Doran, described the state of the art in research on social simulation and defined
social simulation as a particular approach to the study of societies that involved the use
of precisely specified simulation models, often formulated and run on computers. This
approach, said the editors, has been fruitful in the past andappears likely to be even
more so in the future (p. 1). Their forecast has proved to be correct even within the short
time since it was made. Three different indicators suggest that the use of computer
simulations for studying essential aspects of societies has gained importance:
(a) The number of studies relying on this methodology has increased.
(b) The number of disciplines involved has also grown. Once, studies in social simulation
were drawn from the central core of the social sciences and from computer science.
Now other disciplines, such as cognitive science, biology, neuroscience, and some
artificial intelligence (AI) subfields, such as distributed artificial intelligence and
research on multi-agent systems, have been showing a growing interest in the
computer-based study of societies. The motivations that are leading people with a nonsocial scientific background to this common area of investigation are obviously
heterogeneous. However, there are some broad questions of cross-disciplinary interest
that seem to demand the use of computer simulation. These include evolutionary and
dynamic models, the problem of complexity and the paradigm of emergence. Later in
this chapter we shall return to these questions and to the necessity of using computer
simulation to deal with them successfully.
(c) The number of theoretical perspectives involved has increased. Even radically
different approaches, such as conventional AI rule-based systems and research based
on neural nets and genetic algorithms, share the use of computer simulation as their
methodological grounding. The more differentiated the conceptual and theoretical
framework of social simulation becomes, the more influential and significant will be
its scientific role.
Doran and Gilbert (1994) argue that computer simulation is an appropriate methodology
whenever a social phenomenon is not directly accessible, either because it no longer
exists (as in archaeological studies) or because its structure or the effects of its structure,
i.e. its behaviour, are so complex that the observer cannot directly attain a clear picture of
Artificial societies
what is going on (as in some studies of world politics). The simulation is based on a
model constructed by the researcher that is more observable than the target phenomenon
itself. This raises issues immediately about which aspects of the target ought to be
modelled, how the model might be validated and so on. However, these issues are not so
much of an epistemological stumbling block as they might appear. Once the process of
modelling has been accomplished, the model achieves a substantial degree of autonomy.
It is an entity in the world and, as much as any other entity, it is worthy of investigation.
Models are not only necessary instruments for research, they are themselves also
legitimate objects of enquiry. Such artificial societies and their value in theorizing will
be the concern of the first part of this chapter.
In the second part, we will consider an idea that, although central to social simulation,
remains controversial: the idea of emergence. Emergence is also a major concern for
many fields of investigation adjacent to and intertwined with social simulation; for
example, the study of artificial life, biology, and solid state physics. It recurs in many
chapters in this book and was addressed specifically during the second Symposium on
Simulating Societies, held in Siena, Italy, in July 1993, the meeting at which initial
versions of most of the chapters in this book were first presented. We shall argue that
social simulation may help to reformulate the question of emergence in more specific and
empirically relevant terms.
Introduction
That this also gives an indirect account of how things are, in fact, or have been, is only a
secondary aspect of the work.
This alternative role for social simulation deserves attention. We will not be directly
concerned with purely prescriptive studies, although for a variety of reasons the goal of
optimization has had a strong impact on the computer-based study of social processes.
Another, at least equally important, objective in this work is to realize, observe and
experiment with artificial societies in order to improve our knowledge and
understanding, but through exploration, rather than just through description. In this mode
of research, the target is no longer a natural society, but an artificial one, existing only in
the mind of the researcher. A new target (the artificial system) is created with its own
structure (the architecture of the system) and behaviour (the simulation). When a
simulation is run, the system operates in a certain way and displays certain behaviour.
The simulation may either provide a test of the model and its underlying theory, if any, or
may simply allow the experimenter to observe and record the behaviour of the target
system. As the emphasis shifts from describing the behaviour of a target system in order
to understand natural social systems the better to exploit the behaviour of a target for its
own sake, so the objective of the research changes to the observation of and
experimentation with possible social worlds. With the possibility of constructing artificial
systems, a new methodology of scientific inquiry becomes possible.
The value of building artificial societies is not to create new entities for their own
sake. Such an approach, although common to much research and debate about artificial
life, has only moderate scientific interest. It is exemplified by some philosophical
disputes about whether artificial beings may be considered as living (see Emmeche 1994,
Harnad 1994, Langton 1994). These disputes are of little relevance because their
resolution depends entirely on how the notion of life is defined.
Our stress, instead, is on a new experimental methodology consisting of observing
theoretical models performing on some testbed. Such a new methodology could be
defined as exploratory simulation. The exploratory aim synthesizes both the
prescriptive and descriptive objectives: on the one hand, as with the testing of existing
theories, the aim is to increase our knowledge; but on the other, as happens with studies
orientated to the optimization of real life processes, the aim is not to reproduce the social
world, but to create new, although not necessarily better, or more desirable, systems.
Here lies the difference from optimization research.
Exploratory research based on social simulation can contribute typically in any of the
following ways:
(a) implicit but unknown effects can be identified. Computer simulations allow effects
analytically derivable from the model but as yet unforeseen to be detected;
(b) possible alternatives to a performance observed in nature can be found;
(c) the functions of given social phenomena can be carefully observed (we will return to
this issue later); and
(d) sociality, that is, agenthood orientated to other agents, can be modelled explicitly.
In the next section, we shall show why simulation is of particular value for the
development of social theory about sociality.
Artificial societies
Introduction
(a) there are many forms of co-operation and conflict, as well as many forms of altruistic
and selfish, desirable and undesirable social action;
(b) the alternatives are not clear-cut: there are forms of apparently cooperative action that
may be regarded as being conflictual. Social exchange, for example, although
notionally a form of co-operation, may give rise to manipulation and produce
inequality;
(c) many concrete examples of sociality are neither advantageous nor harmful to cooperation. For example, conversation is itself neither intrinsically conflictual nor cooperative. Furthermore, many abstract categories of social action are neither pro- nor
anti-social. Consider, for example, communication, which may be either co-operative
or aggressive; and influence and persuasion, which may be either selfish or altruistic;
and
(d) a social action may be governed by a complex hierarchy of goals so that a given
actions lower-level goals may differ from the higher-level ones. A pro-social action
may be included in an anti-social plan and vice versa. Deception may be used for
helping, and co-operation for cheating.
Natural interaction is undoubtedly much richer than is allowed by the basic categories of
co-operation and conflict. To understand and model social action and account for the
complex and varied forms of interaction means working out subtler categories of analysis
than are usually employed by social thinkers. The fine-grained modelling of sociality
allowed by the study of artificial society could greatly enhance our capacity to account
for the complexity of interaction.
Self-sufficient agents and social interference
Often, social agents are thought of as self-sufficient beings acting in a common world.
The term applied by some social scientists to describe structural social relations is
interference, essentially referring to agents accidentally hindering or facilitating one
another. This view is limiting, since it does not account for the role of the structures that
pre-exist interaction (Conte & Sichman, in press). It does not show that certain forms of
social action are present in nuce in certain types of structural relations and emerge from
them. A detailed description of different types of social structures would allow the
relationships between social structures and social action to be modelled in more revealing
ways.
Groups and coalition formation
The classical view of groups and coalition formation is based on two ideas: first, coordination to overcome interference and, secondly, collaboration involving social
intentions and beliefs.
Co-ordination
If agents are conceived to be self-sufficient beings, once they are in a social context they
will find constraints to their self-fulfilment. The common social environment and the
Artificial societies
other agents limit their autonomy and achievements. This may be ameliorated through
co-ordination of the agents actions.
Collaboration
Co-ordination alone is insufficient for any sort of coalition or agreement to take place
among rational agents. Agents also need to have some degree of awareness of others
existence, wants and habits. In other words, agents need to have some mental
representation of other agents minds. These representations can be divided into two main
categories:
(a) Strategic beliefs (in the sense of strategic rationality), which consist of individuals
mental states produced by, and taking into account, the mental states of other, possibly
interfering, agents.
(b) Group beliefs and intentions (Gilbert 1987, 1989; Tuomela 1991, 1992), which
consist of the sharing of similar beliefs and intentions by an aggregate of agents.
Social thinkers have been debating what exactly should be meant by we-ness and
we-intentions, and where precisely the core of groups and teams resides, without
reaching any satisfactory agreement.
However, these two ingredients are not in themselves sufficient to account for coalition
formation. They do not allow the gap between the individual level of agency and the
group or collective level to be filled. Or, when some attempt in this direction is made, as
happened with the notion of strategic knowledge provided by game theorists, only a
simplistic, reductionist view of social mental states could be obtained. What is lacking is
a fine-grained modelling of sociality that allows for the variety of social beliefs and goals
to be taken into account. Agency should not be seen as a twofold, individual and societal,
phenomenon, but as a multi-level one, where individual and societal levels are integrated
because of the special make up of the agentstheir social characterization. It is this level
of sociality that throws a bridge between individual and collective action.
In sum, the fine-grained modelling of sociality is extremely important, both as an
independent objective of scientific inquiry and as an instrument and a condition for the
study of societies. Since sociality represents the medium between individuals and
collectives, the study of various forms and levels of social organization cannot do without
the study of sociality. However, modelling of sociality requires:
(a) detailed and explicit models of social agents;
(b) an observatory of various forms and levels of sociality, not predetermined and
constrained by existing theories and assumptions;
(c) the possibility of experimenting with the observed social phenomena; and
(d) the possibility of observing the emergent properties of the phenomena.
None of these requirements can easily be fulfilled without computational models of social
agents and without computer simulations of social processes.
Introduction
Artificial societies
Sub-cognitive bias
The idea of emergence has most often been applied to systems with subcognitive units
(for example, reactive systems in AI and subsymbolic systems working on a neural
network base). Emergence is meant to have the same role for the subcognitive as
knowledge and calculation have for the cognitive. But emergence can also be significant
among collections of cognitive agents. The following effects can emerge from interacting
cognitive agents (see also Conte & Castelfranchi in press):
(a) the emergence of a structure of objective relationships among a set of unaware
agents from their individual goals and abilities once agents are placed in a common
world;
(b) the emergence of awareness by agents of these precognitive relationships because of,
for example, adaptive cognition (learning from failures);
(c) the evolutionary emergence of internal goals, complex representations and minds
from environmental pressures (fitness) exerted on behavioural systems;
(d) the emergence of collective activity, teamwork and group intentions from
complementarity of goals and actions, sharing of beliefs, etc.; and
(e) the emergence of co-operation from the functional effects of the agents deliberate
actions without the agents becoming aware of it.
The behavioural bias
Even when applied to rational agents, it is usually only at the behavioural level (the
spreading of given behavioural patterns and strategies, such as co-operative or defective
choices, the learning of seemingly planful routines in fundamentally reactive systems,
etc.) that behavioural properties are noted; non-behavioural emergent effects (e.g.
cognitive structures) are usually ignored. For example, the idea that cognitive structures
(the capacity for knowledge-based reasoning, planning, decision-making, etc.) represent a
highly adaptive emergent response developed by complex systems under environmental
pressure has not received the attention it deserves.
The individualistic bias
The study of emergent social phenomena tends to imply an individualistic approach,
tacitly adopting one of the traditional solutions to the micro-macro problem, namely that
phenomena at the macro-level can be said to emerge from phenomena at the micro-level
of analysis and interpretation. The micro-macro link is therefore reduced to one-way
(from micro to macro) relationships. However, individual social action can be forged by
macro-social forces of various sorts, for example, social norms. Of course, one may claim
that a norm is only a convenient term to describe a state of affairs that has no real
correspondence in the minds and actions of social agents. Alternatively, one may claim
Introduction
that social norms correspond to specific phenomena that are produced spontaneously and
gradually by interactional practice. Then social norms (macro effects) tend to constrain
not only the actions of agents, but also their minds and beliefs, their goals and even their
evaluations, expectations and attitudes (a micro effect). The formation of new mental
constructs can be a by-product of such norms. Over time, these mental constructs tend to
form part of the social characterization of the agents. What are they but emerging microproperties of macro-social phenomena?
Reconsidering emergence
A unifying and better-defined notion of emergence could be obtained if the above biases
were avoided. A unifying notion of emergence cannot decide to account only for
properties of subcognitive systems in interaction. Nor can emergent properties be reduced
to observable, behavioural effects (leaving aside internal, or mental, structures) at the
micro-level, ignoring the fact that actions at the micro-level can be considered to be the
emergent effects of phenomena at the macro-level.
It is also necessary to find criteria that allow arbitrary, accidental or irrelevant
emergent effects to be discriminated from relevant ones. This bring us back to previous
functionalist debates. The puzzles left unsolved by structural-functionalist theory may be
concealed under the apparently new terminology of emergence. However, the
methodology of computer simulation offers an opportunity for the social sciences to
address some of these puzzles. Within such a methodology, a functional interpretation of
a given phenomenon can be tested and falsified, since the three main questions that
should be addressed within a functional explanation: (a) what the limits are within which
a given effect is functional; (b) who the beneficiaries of the functional effects are; and (c)
what the mechanisms of reproduction and selection are (cf. Elster 1982), through which a
given effect becomes functional, can finally be answered by means of computer
simulation.
To summarize, the importance of emergence is that interacting agents in a common
world produce phenomena that are not necessarily intended by the agents themselves but
are none the less relevant to their later behaviour and achievements. Social simulation
represents a testbed for the study of emergence. A scientifically adequate exploration of
emerging phenomena is not practicable in the absence of such a methodology.
Artificial societies
The role of simulation in evaluating models of both natural and artificial societies and the
importance of the concept of emergence will become apparent in the remaining chapters
of this book. The chapters in the first section, on the simulation of social theories, focus
on building models that articulate social theories in areas of the social sciences which
previously have been the preserve of more traditional methods of analysis. For example,
in Chapter 2, Robert Axelrod examines how political actors can emerge from
aggregations of smaller political units. He points out that the major research paradigm for
formal models in politics has been game theory, but game theory takes the actors
involved as a given. He shows how clusters of actors that behave like independent
Artificial societies
10
political states emerge from the interactions of basic units during the simulation of a
tribute model in which the basic units compete with each other for wealth.
Massimo Egidi and Luigi Marengo, in Chapter 3, also use simulation to shed new light
on a topic on which much has previously been written: the division of labour in
organizations. They model a simple production process: the counting of a heap of
banknotes by a group of agents. Each agent is represented by a form of program called a
classifier system which is able to evolve the rules it uses to determine how many
banknotes it will count and where it will pass on the result. They test the simulation in a
variety of conditions and show that while a centralized and hierarchical division of labour
is most efficient when the agents are infallible, if the agents make mistakes in counting, a
decentralized control mechanism that permits local adjustments is better.
In Chapter 4, Jean Pierre Treuil applies social simulation to a problem of social
anthropology. He develops a model of the emergence and transmission of kinship
structures in Australian aboriginal societies, using simulation as an alternative to the
more usual mathematical models, in order to show the conditions under which the kinship
system converges to an equilibrium pattern of unions. In Chapter 5, Stphane Bura and
her colleagues describe a simulation located within the conceptual framework of central
place theory that aims to account for the location of human settlements. They argue that
simulation has advantages over other types of modelling, such as using differential
equations, because it allows a much greater variety of qualitative and quantitative factors
to be brought into play. Their model is used to study the genesis, development and
concentration of urban functions during the long-term evolution of settlement systems.
They show that despite the highly varied local conditions in which settlements are
located, there is a remarkable convergence in the pattern of settlements in the system as a
whole.
While all the previous chapters have reported simulations in which the agents are
intentionally very simple, having at most a few rules to determine their interactions with
other agents, Jim Doran and Mike Palmer in Chapter 6 describe a simulation designed to
explore the growth of social complexity in societies in the Upper Palaeolithic period in
south-western Europe in which the individual agents are comparatively complex. This is
because the agents cognitive representations are modelled explicitly. The authors argue
that it is essential to model humans unique cognitive abilities and show that, by doing so,
their simulation illustrates the emergence of leaders, these being the agents that seem to
offer the best plans for action to the others.
In the final chapter of this section, Roger McCain suggests that it is possible to model
consumer demand in microeconomics by using simulations based on genetic algorithms.
He investigates whether the agents eventually learn optimal behaviour, that is, whether
they are able to maximize their utility, thus tackling from a new perspective one of the
foundational problems of microeconomics.
In the second section of the book, the contributors consider the idea of emergence
from a number of different perspectives. Nigel Gilbert in Chapter 8 reviews the debate in
sociology between methodological individualism, the idea that macro phenomena must
be accounted for entirely by the actions of individuals, and methodological holism: the
idea that individuals behaviour is to be entirely explained in terms of their locations in
society. He suggests that a more recent position, structuration, is more illuminating
than either of these extreme positions. Comparing the theories underlying most
Introduction
11
simulation models of societies with structuration theory, he notes that so far simulations
have failed to take account of the fact that people have the ability to discover and monitor
the emergent features of their own societies. He argues that this ability has major
implications for human societies and needs to be considered in developing future models.
Edwin Hutchins and Brian Hazlehurst in Chapter 9 describe a simulation that
illuminates the emergence of natural language. They point out that for people to
communicate, there needs to be a shared lexicon in which each item has an agreed form
and an agreed meaning. The puzzle is how this lexicon could have evolved in human
societies from nothing. They develop an artificial society consisting of a number of
neural nets, warrant it in terms of a theory of cognition and then present a simulation in
which the interacting neural nets evolve a common lexicon for describing visual scenes.
The simulation is even able to reproduce the emergence of dialects among some
subgroups of nets.
In Chapter 10, Alexis Drogoul and his colleagues describe the work they have been
doing on the simulation of the social organization of an ant colony. They report the
results of their simulation of sociogenesis, that is, the creation of a society from one
individual, the queen, and compare the macro-level, emergent behaviour of an artificial
ant nest with records of the growth of natural nests, showing that the two are similar. In a
further set of experiments with their artificial nests, they simulate the effect of having
several queens and observe that although the queens are programmed to be in
competition with each other for resources, the resulting behaviour can easily be
interpreted in terms of co-operation and altruism. They conclude that many behaviours
viewed as being co-operative can be obtained by competitive interplay between agents.
Nicholas Findler and Raphael Malyanker in Chapter 11 return to the theme of alliance
formation considered by Axelrod in Chapter 2, but from a different perspective. Their
emphasis is on the development of norms and in particular on the mechanisms by which
norms evolve and are im-plemented in international power politics. They use the
techniques of distributed artificial intelligence (which is also the starting point of the
work of Doran and Palmer, and Drogoul et al.) to model agents that can reason about
their environment and interact with each other.
In the third and final section are three chapters that address some fundamental issues
in social simulation. In all three chapters, the principal concern is not to model some
natural society or even an approximation to a natural society, but to develop methods and
ideas that can be applied to both artificial and natural societies. In Chapter 12, Domenico
Parisi et al. describe simulations of the evolution of altruism and attachment behaviour in
a population of neural networks and show that it is possible to observe the emergence of a
type of behaviour that tends to keep individuals close together, thus creating what might
be described as a simple social group. Rosaria Conte and Castelfranchi in Chapter 13
examine the role of norms in controlling aggression among competing agents that have
some limited ability to reason. They carry out a series of simulation experiments on a
simple artificial society in which there is competition for food, to compare the effects
of individual rational strategies and societal norms. They find normative controls result in
a more equitable distribution of resources in the population.
Social simulations have usually been programmed using either conventional computer
languages such as C and Basic or languages developed for artificial intelligence research,
such as LISP, Prolog and Smalltalk. In the final chapter, Michael Fisher and Michael
Artificial societies
12
Introduction
13
Part I
The simulation of social
theories
Chapter 2
A model of the emergence of new political
actors
Robert Axelrod
How can new political actors emerge from an aggregation of smaller political actors?
This chapter presents a simulation model that provides one answer. In its broadest
perspective, the work can be seen as part of the study of emergent organization through
bottom-up processes. In such bottom-up processes, small units interact according to
locally defined rules, and the result is emergent properties of the system such as the
formation of new levels of organization. Thus the work is typical of the Santa Fe
approach to complex adaptive systems (Stein 1989; Fontana 1991; Holland 1992). The
concern with increased levels of organization is also reminiscent of how biological
systems succeeded in making the transition from single-celled organisms to multiplecelled organisms (Buss 1987), and how brains function by organizing individual neurons
into meaningful structures (Hebb 1949; Minsky 1985).
The task at hand involves the emergence of new political actors. This is a vital
question in the post-Cold War world. We are experiencing an era in which the standard
unit of politics, the nation, is no longer completely stable. We see, on the one hand, that
some states are disintegrating, as in the former Soviet Union and Yugoslavia. We see, on
the other hand, that larger units are being organized, such as the European Union (EU),
and other regional associations. The question of the aggregation and disaggregation of
political actors is essential for the understanding of the future of global politics, both in
terms of international security affairs and international political economy.
The question of how the world can be placed on a sustainable path of development is a
particularly pressing question. The emergence of new political actors is fundamental to
the question of sustainability. One of the main problems of attaining sustainability is the
tragedy of the commons (Hardin 1968). The tragedy of the commons arises when many
independent actors (people, villages, states, or whatever) each over-graze because there
is no mechanism to enforce the collective interests of all against the private interests of
each. This leads to resource depletion, elimination of bio-diversity, overpopulation, war,
and other major social problems. A major route to the prevention of the tragedy of the
commons is the emergence of a political actor based upon the organization of previously
independent actors. Today we have political actors at the national level that can regulate
resource use within their boundaries, but we do not yet have very effective political
actors at the transnational level to regulate resource use at the global level.1
Political scientists have access to a variety of concepts and theories to analyze the
emergence of new political actors. Unfortunately, they do not have any formal models
that account for this emergence endogenously. In fact, the problem is much like
Artificial societies
16
17
Coercion in general and extortion in particular have played a central role in the
formation of states over the centuries (Tilly 1985; 1990). For this reason, at the heart of
the present model is the simple dynamic of pay or else. An elementary actor can make
a demand of a neighbour for payment of resources, with the threat that if payment is not
forthcoming there will be a war.
In the tribute model wars result in changes in wealth (and thus power), but not in
outright territorial conquest.2 This allows each territory to maintain itself as a separate
actor, thereby allowing for the possible emergence of sets of territorial actors who might
form a new aggregate actor. Although no territory changes hands, war results in costs to
both sides, but especially to the weaker side. Thus the heart of the model is a tribute
system in which an actor can extract resources from others through tribute payments, and
use these resources to extract still more resources. Alliances are also allowed so that
actors can work together. Whether a set of elementary actors emerges as a stable
aggregate actor will depend on the dynam-ics of the tribute system: whether groups of
actors emerge which function as a single aggregate, and whether alliance patterns emerge
which lead to stable, coordinated actions.
Unlike my earlier work on the Prisoners Dilemma (e.g. Axelrod 1984) the tribute
model is based upon extortion rather than co-operation. Also, unlike the earlier work, it
does not assume that the actors are equal in power, but instead the tribute model takes
power differences as vital, and as resulting directly from the dynamics of the model. Nor
does it assume that the actors interact only two at a time. Like my earlier work, it is not a
rational model. The tribute model assumes that actors develop more or less strong
commitments to each other based upon their prior actions. These commitments can be
thought of as the result of psychological processes (e.g. to stand by those who have
helped you in the past), or the result of political rules of thumb (e.g. to support those who
may later help you in your time of need). Actions are based upon simple decision rules
rather than game theoretic calculations of optimal choice, since rational calculations
would be virtually impossible to make in such a complex setting. An important and novel
feature is that the behaviour of the actors changes over time as they apply simple decision
rules to data about the historical experience they have gathered through prior interactions.
The project will be successful to the extent that it can account for important
phenomena that cannot be accounted for by existing formal theories, and can do so with
only a few assumptions. The main result to be generated is the emergence of new
political actors at a higher level of organization, and the main assumptions have to do
with how elementary actors interact with each other. As we shall see, the model generates
spontaneously not only emergent actors, but also other interesting phenomena that exist
in the world of international politics.
Artificial societies
18
between actors on the ends of the line and those in the interior, the line is joined into a
circle. This gives each of the ten actors only two neighbours.
There is one resource in the model, called wealth. Each actor is given some initial
wealth. The initial wealth is chosen from a uniform distribution between 300 and 500.
These parameters, like all the others in the model, are somewhat arbitrary and are
selected for convenience.
The basic cycle of the model is called a year. In each year, three actors are chosen
one after another at random to become active. An active actor, A, may demand tribute
from one of the other actors, B. Initially, the target, B, must be a neighbour, but later this
restriction will be relaxed when alliances are considered. The model is based upon a
dynamic of pay or else. The target, B, then has a choice of paying tribute to the
demander, or fighting.
The selection of actors to be active is based upon the notion that ambitious leaders and
potential disputes arise at random. But once given an opportunity to make a demand, the
activated actor need not actually make a demand if it finds that the current situation is not
favourable.
If A does make a demand, the target B has a choice.
If B pays, wealth is transferred directly from B to A. The amount of wealth transferred
to A is 250 if B has that much. Otherwise, the amount transferred to A is whatever B
has. This transfer represents tribute that B pays to A to avoid a fight.3
If B fights rather than pays, each side loses 25 per cent of the other sides wealth (or
both lose proportionally less if either side doesnt have that much wealth).4 This is a
simple Lanchester attrition dynamic (Lanchester 1916; Epstein 1985). The idea is that
in a fight both sides suffer, but the stronger side imposes more damage than the
weaker side does.5
After the three activations, the yearly cycle ends with a harvest which increases each
actors wealth by 20. This feeds some new wealth into the system. A typical run is 1000
years.
The next question to consider is decision rules used by the actors for making and
responding to demands. It would be almost impossible for actors to develop fully rational
rules in the tribute game because of its complexity over time and space. Therefore, the
model uses heuristic decision rules that capture some of the key short-term considerations
facing the players.
The active actor needs to decide of whom, if anyone, to make a demand. The ideal
target of a demand is weak enough so that it might choose to pay rather than fight, and
so that it will not cause much damage if it does choose to fight. On the other hand, the
ideal target should be strong enough to be able to afford to pay as much possible. A
suitable decision rule combining both of these considerations is to choose among the
potential targets the one that maximizes the product of the targets vulnerability
multiplied by its possible payment. The targets vulnerability is (WAWT)/WA where
WA and WT are the wealths of the active actor and the target, respectively. The targets
payment is how much the other side can pay, which is the maximum of its wealth and
the full tribute of 250. If no potential target has positive vulnerability (i.e. no potential
target is weaker than the demander), then no demand is made.
19
The decision rule used for the target is simpler: fight if and only if it would cause less
damage than paying would.
So far, the model has considered only pairwise interactions. But in order to study the
development of emergent actors, there has to be a way for the basic actors to work
together. The key idea is that actors develop degrees of commitment to each other. These
commitments are caused by their choices to pay or fight, and in turn have consequences
for how they will pay or fight in the future. The basic idea is that if two elementary actors
fight, another adjacent actor will join the side to which it has greater commitment. If it
has equal commitment to the demander and the target, it stays neutral. If it does join one
side or the other, it contributes forces (i.e. wealth) in proportion to its commitment to that
side.
Initially, no one has any commitments to others, and each actor is fully committed to
itself. Commitment of actor i to actor j increases when:
(a) i pays tribute to j (subservience);
(b) i receives tribute from j (protection); or
(c) i fights on the same side as j (friendship).
Similarly, commitment decreases whenever:
(d) i fights on the opposite side to j (hostility).
The motivation for these rules of commitment dynamics is simple. When one actor pays
tribute to another, the first actor is also likely to be partially under the second actors
political domination, and therefore compelled to assist the second actor next time. A state
typically becomes committed to helping the patron, whether by choice or necessity, as
illustrated by the participation of many Latin-American states in the Second World War
after Pearl Harbor. Conversely, the protection of states that have provided benefits is also
commonplace, to protect future sources of revenue. The next point is that if two sides
have fought together in the past, they tend partially to be committed to each other for the
future, as in the case of the United States and South Korea after the Korean War. On the
other hand, two states that have fought each other are less likely to support one another in
the future, as in the case of Germany and France after the First World War.6
The model assumes that increases and decreases of commitment are in constant
amounts, namely increments of 10 per cent. In addition, one actors commitment to
another can never be more than 100 per cent nor less than 0 per cent. The commitment
processes described above maintain the symmetry of commitment. This is because two
actors start with no commitment to each other and their commitments to each other
always grow and decline in unison. For example, if two actors fight on the same side in a
war, then their commitment to each other will increase by 10 per cent.
The final part of the model deals with co-ordination of actors. To keep interactions
similar to land combat, co-ordinated action is assumed to require contiguity. Thus an
actor is an eligible target for a demander only if everyone between them joins the
demander. Others may then join the demander or the target provided that the intermediate
actors also join in: for example, the actor in position 5 may make a demand on actor 8
only if actors 6 and 7 join actor 5. This requires that actors 6 and 7 are both more
committed to 5 than to 8. Then, if 5 does make a demand on 8, actor 10 may join the
defence of 8 only if 9 joins too.
Artificial societies
20
Commitments and wealths are common knowledge.7 Thus when an active actor is
evaluating the vulnerability of an eligible target, it can take into account the commitments
and wealth of all actors who would join either side. Similarly, the target of a demand can
determine the cost to itself of fighting, by calculating the damage that the attacking
alliance could do, and the proportion of that damage that the target would suffer, which is
the proportion of the defending alliances wealth contributed by the defender.
21
Figure 2.1 shows the wealth of each of the ten actors over the long timespan of 1000
years. The figure clearly shows that three actors were
Artificial societies
22
23
Artificial societies
24
i, j
10
100
30
10
40
30
100
20
20
100
20
20
100
30
30
100
100
100
50
100
50
10
20
50
100
10
10
40
20
10
100
The other common pattern of commitments develops after this, and consists of two
clusters of actors. A cluster can be defined as a set of actors, all of whom are highly
committed to each other (say at the 50 per cent level). Table 2.2 shows the same
population 25 years later, divided neatly
10
100
100
70
100
100
100
100
100
100
100
100
100
70
100
100
100
90
70
100
100
60
100
100
100
100
100
60
100
100
100
100
100
100
100
100
100
100
100
100
100
100
100
90
100
100
100
10
100
100
70
100
100
100
25
into two clusters, with no commitments at all between members of different clusters.9
To see how this population can become divided so quickly into two distinct clusters,
look at the sequence of payments and fights involving a single illustrative actor, actor 5.
Table 2.3 shows the fights and payments involving actor 5 from year 25, when it had only
one partial commitment, to year 50, when it was fully integrated into a cluster of four
actors.
The events of year 35 can serve to illustrate the dynamics of commitments that lead to
the development of distinct clusters. In that year, 3 and 4 fought 5, 6 and 7. This led 5 to
increase its commitment to 6 and 7 while decreasing its commitment (if any) to 3 and 4.
In general, as fights take place, they increase the commitments of those fighting on the
same side, and decrease commitments of those fighting on opposite sides. Thus, as
clusters begin to form, then tend to get even more distinct. This is because pairs of actors
who fought together became even more committed to each other, and pairs on opposite
sides lost whatever partial commitment they may have had to each other. By year 45,
fights were taking place that involved all ten actors, leading to the pattern of two strong
and distinct clusters that was shown in Table 2.2
Having seen how commitments can lead to clusters in an illustration from population
2, let us now return to the unusual case of population 1, where the strongest actor
collapsed. The dynamics over the full 1000 years was shown in Figures 2.1 and 2.2. Now
we can look in detail at the period of collapse that took place between years 890 and 950.
This is shown in Figure 2.7.
In year 911, actor 10 targeted actor 9, resulting in a world war in-
Active
actor
Target
actor
Roles
30
5 aAD
30
5 PR
32
4 ADd
32
5 aADd
3, 4
33
3 dDAa
2, 3
34
7 RP
35
4 dDAaa
6, 7
3, 4
36
4 aADd
2, 3
36
aaddDAaa
6,7
1, 2, 810
37
5 PR
38
aaaddDAaa
3, 4
6,7
13, 810
Artificial societies
26
38
aAaddDaaa
6, 7
13, 810
41
aaaddDAaa
6,7
13, 810
42
5 PR
42
4 PR
44
7 RP
46
dddaaaADdd
4, 6, 7
13, 810
47
dddaaaADdd
4, 6, 7
13, 810
48
ddDaaaAddd
4, 6, 7
13, 810
48
ddDaAaaddd
4, 6, 7
13, 810
Key to roles: A=attacker; a=attackers ally; D=defender; d=defenders ally; P=payer; R=receiver of
tribute.
a) Increases by 10%, up to 100%; b) Decreases by 10%, down to 0%.
27
more damage was done. The result was a sharp decline in the wealths of the largest actors
(5 on one side and 2 and 4 on the other).
This dramatic decline of the strongest actor is because it was dragged into fights
involving weak actors to whom it had developed commitments. This is reminiscent of
what happened to two dominant powers, the Hapsburgs in the seventeenth century, and
Britain in the nineteenth to twentieth centuries. Paul Kennedy has an apt term for the
dangers of the excess commitments that strong powers tend to take on. It is imperial
overstretch (Kennedy 1987). This is just what is illustrated in Figure 2.7. It is striking
that a simple model developed to generate new political actors produced behaviour that
has been the subject of a major policy debate in contemporary international affairs (e.g.
Nye 1990).
Clusters are not only settings for mutual commitment; they are also settings for
extraction of tribute. The strongest actor in a cluster typically stays strong and grows by
making demands and collecting tribute from the other members of its cluster. As actors
slowly grow due to their annual harvest, the strong member makes demands upon the
weak and reduces their wealth. In fact, as the strong member selects lucrative targets, its
demands tend automatically to be rotated among members of the cluster. This rotation
has the unplanned consequence of preventing any target from growing very strong. But
sometimes a weak member of a cluster escapes demands just long enough to grow strong
enough to make its own demands on still weaker members. If luck lets it be active often
enough at this critical stage, it can collect sufficient tribute so that the strongest member
no longer finds it advantageous to challenge the newly-wealthy actor. An example of this
process occurred after year 100 in population 3 (see Figure 2.8). In this example, actor 5
has managed to grow in the shadow of
Artificial societies
28
Demands within clusters do not always get resolved by payments for tribute. In fact,
civil wars can occur, i.e. fighting among members of the same cluster. The strongest
member of a cluster typically is strong enough, by taking sides, to prevent such a fight
from taking place. If, however, the strongest member is equally committed to the attacker
and the defender (perhaps by being totally committed to both), then it would stay neutral.
This allows civil wars to occur if the active actor finds a lucrative target within the same
cluster, and the target finds it more costly to pay tribute than to fight. The attacker and
defender may even find allies among the others if these others are not equally committed
to the two sides.
Surprisingly, initial differences in wealth do not matter for wealth in the long run. In
five populations of the model there were 50 actors in all, 14 of whom were quite
successful, using a criterion of wealth of over 10,000 after 1000 years. Of these 14
successful actors, half had less than average initial wealth and half had more than average
initial wealth. One might expect that initial wealth would be a great advantage, since it
would allow an actor to make successful demands on its neighbours at first, and thereby
build up a dominant position in a cluster. But having substantial initial wealth can also
make an actor a lucrative target for other strong actors. In addition, having substantial
initial wealth can make one over-confident, given the limited rationality of the decision
rules in effect.10
The results of the models performance can now be summarized in terms of six
characteristics:
1. Things usually dont settle down. Instead, the history of the model shows considerable
complexity. For example, as late as year 900, one of the populations suffered a series
of fights that destroyed over three-quarters of the global wealth (see Figure 2.2).
2. Histories show considerable variability. The combinations of wealthy actors,
frequency of fights, and trends in population wealth differ considerably from one
population to another. Even though each actor is using the same decision rules, the
results differ greatly because of random differences in initial wealth and in the order in
which actors become active.
3.Imperial overstretch can bring down even the strongest actor. As an actor becomes
committed to others via tribute relationships and fighting together, it becomes exposed
to the risks of whatever fights the others get into. Since actors decide on demands and
responses based upon their own calculations, even a weak actor can choose to make or
resist a demand for its own reasons, and thereby drag into the struggle a strong actor
who is committed to it.
4. Civil wars can occur among the smaller members of a cluster. While the strongest
member of a cluster typically can prevent a fight among members of the same cluster
by taking sides, it would not do so if it had equal commitment to the two sides.
Therefore, smaller members of a cluster may fight each other while the strongest
member stands aside.
5. A cluster can have more than one powerful member. Clusters are often like empires,
with one powerful actor and many weaker ones who pay tribute to it. But as we have
seen in the case of Figure 2.8, it is also possible for a second actor to grow strong in
the shadow of the first.
6. Initial endowment does not guarantee or even predict success. Before clusters of
commitments are established, wealth can be as much a handicap as an asset. The
29
reason is that wealth makes an actor a lucrative target for other strong actors, and can
make one over-confident in making demands.
Artificial societies
30
One of the advantages of a simulation model is that it is relatively easy to answer what
if questions about the effects of varying the parameters or the premises of the model.
Here is a very brief sample of some of these questions, with answers based on further
simulation runs.
1. Does the model settle down after 1000 years?
No. Even in runs up to 10,000 years there are repeated large wars, and strong actors
continue to rise and fall.
2. Are there better decision rules?
Yes. An example is adding a constraint so that a demand will not be made of a target
that would find it cheaper to fight than to pay. An individual using this revised rule
does much better than the others. Moreover, if everyone is using the revised rule,
any single individual who uses the old rule would do worse than the others.
3. Does it pay to give and receive commitments?
Yes. An actor who neither gives nor receives commitment does very poorly.
Moreover, if just one actor gives and receives commitments (and the others do not,
except with that actor), then that actor does very well.
4. What happens if everyone can reach everyone else (as in sea power)?
In this extreme case, a single powerful actor tends to dominate, and there are
relatively dense patterns of commitment. Presumably, in a two-dimensional space,
where everyone has more neighbours than in a one-dimensional space but fewer
than in the sea power case, there might be an intermediate level of dominance
and density of commitment.
5. What if wealth does not enter into anyones calculations?
If the demands are made without regard to wealth, and if everyone always fights,
then the result is two clear clusters, one of which has virtually all the wealth.
6. Do islands of commitment grow?
Yes. If two adjacent actors initially have just 10 per cent commitment to each other
(and all other commitments are zero), then these two actors develop complete
commitment to each other, and both tend to prosper.
31
and presenting data that help us to see the forest for the trees. This chapter illustrates four
different kinds of historical analy sis. In effect, these are four different ways of looking at
history.
1. History from the point of view of a single actor: plots of relative wealth of individual
actors over time.
2. History from a global point of view: joint plots of wealth and number of fights over
time.
3. History from the point of view of the emergence of new actors: commitment matrices
revealing emerging clusters of actors.
4. History as news: event data, including chronologies of tribute payments and wars
fought between alliances.
Perhaps the most useful outcome of a simulation model is to provide new ways of
thinking about old problems. In the present case, the need to determine whether or not the
model was successful in generating new political actors forced the specification of
explicit criteria for recognizing a new political actor should one arise. This in itself is a
useful exercise. In addition, the same need led to the development of different ways of
viewing history, and especially to ways of analyzing the interaction among emerging
clusters of actors.
In the future it would be good to use these conceptual and statistical developments to
answer some new questions suggested by the model. For example, the dynamics we have
seen in the tribute model suggest the following interesting questions:
1. What are the minimal conditions for a new actor to emerge?
2. What tends to promote such emergence?
3. How are the dynamics affected by the number of elementary actors?
4. What can lead to collapse of an aggregate actor?
5. How can new actors grow in the shadow of established actors?
While no one would believe that the answers to these questions given by the model
itself would necessarily be accurate for the real world, a realistic hope is that the concepts
developed in generating the answers for the model would be useful in providing new
ways of thinking about comparable questions in the real world. Indeed, understanding the
dynamics of the tribute model might also suggest specific hypotheses that could be tested
with data from the real world. In addition, the study of one model (such as the tribute
model) can provide the basis for insights into how even simpler models can be developed
that might allow deeper insights into issues of emergent political organization.
Finally, a simulation model such as the tribute model can lead to insights into where
there might be policy leverage in the real world. For example, one might be able to
identify situations in which slight changes in the system could lead to substantially
improved outcomesin terms of fewer destructive wars or greater co-ordination of
political action. If we knew when minor inputs could lead to major gains we would have
Artificial societies
32
valuable clues to consider for policy leverage in the real worldthe world where
avoiding war and achieving sustainability really matter.
Acknowledgements
I thank Rob Axtell, Arthur Burks, Lars-Erik Cederman, Michael Cohen, Paul Courant,
Joshua Epstein, Walter Fontana, Stephanie Forrest, Murray Gell-Mann, Nigel Gilbert,
John Holland, John Miller, Melanie Mitchell, Robert Pahre, Rick Riolo, Carl Simon, and
Gerard Weisbuch. For financial support I thank Project 2050 through the Santa Fe
Institute, the National Science Foundation, and the LS&A Enrichment Fund of the
University of Michigan.
Notes
1. Of course, the ability to regulate resource use does not guarantee that the ability will be
wisely used. For example, the former Soviet Union had the power to control pollution from
its factories, but in the interests of maximizing production it chose not to exercise that
power.
2. This contrasts with Cusack & Stoll (1990), whose simulation model uses conquest of territory
as the major dynamic.
3. Having a fixed maximum demand is arbitrary, but avoids the need for the actors to calculate
what demand they will make.
4. For example, if A has 400 and B has 300, A loses 0.25*300=75, and B loses 0.25*400=100.
The wealths after the fight would then be 40075=325 for A, and 300100=200 for B. Note
that the disparity in wealth increases from 100 to 125. If B began the fight with only 50, then
B would lose all 50 (which is half the other damage A is capable of doing), and A would lose
half the maximum damage B is capable of causing, i.e. 0.25*50*50=12.5.
5. The demander is assumed to carry out its implicit threat to fight if the demand is not met.
This assumes, in effect, that the demanders need to maintain its reputation is strong enough
to maintain the credibility of the threat.
6. For evidence on the friendship and hostility aspects, see Axelrod & Bennett (1993).
7. In international politics, wealth can usually be estimated with rough accuracy, but
commitments are harder to predict. There is a rich empirical and theoretical literature on
commitment. See, for example, Schelling (1960), Huth (1988), Powell (1990), Bueno de
Mesquita & Lalman (1992) and Axelrod & Bennett (1993).
8. A run takes only 23 seconds on a Macintosh Quadra 700 using Pascal. Of course, the analysis
can take days or weeks.
9. In many cases, there were two large clusters, with one or two actors participating in both the
clusters.
10. For example, suppose an actor with wealth of 500 makes a demand on a neighbour with 400.
The neighbour will fight rather than pay (losing 0.25*500=125, which is less than payment
of the standard demand of 250). Since the target fought, the demander loses 0.25*400=100,
and is worse off by having made the demand. Had the initial wealths been greater than 1000,
this would not happen, since the cost of fighting would be greater than the standard demand
of 250.
Chapter 3
Division of labour and social co-ordination
modes: a simple simulation model
Downloaded by [Universidad Autnoma de Madrid] at 04:28 05 November 2013
It is customary to consider economic organizations as social systems which make the coordination of individual plans and decisions possible. However, little attention has been
given to the connection between forms of coordination and the process which
characterizes industrial societies: the process of increasing division of labour.
The fact that co-ordination and the division of labour have been somewhat connected
in their historical development is hardly possible to doubt: as economic organizations
increased their degree of division of labour and knowledge, the problems of co-ordination
among a growing number of increasingly inter-related producers and decision-makers
became more and more complex. Co-ordination of distributed decisions by markets,
firms and other economic institutions appears as the other side of the process of the
increasing division of labour.
The historical evidence does not show clearly whether the two proc-esses of coordination and the division of labour have co-evolved as two aspects of the same
phenomenon or have in the main proceeded independently, in terms of both temporal and
causal relationships, but some evidence at the microeconomic levelsuch as analyses of
the processes of design and managerial planning in modern corporationsseem to
support the hypothesis of co-evolution. In economic organizations in which planning and
design are highly purposeful activities, division of labour and co-ordination are the joint
result of these very activities. But can the hypothesis of co-evolution of co-ordination and
division of labour be extended also to the cases in whichboth in markets and in
organizationsthey are not the outcome of a purposeful planning and design process, but
Artificial societies
34
are emergent and partly unintended properties of the interaction between distributed
decision-making activities (cf. von Hayek 1948)?
Behind this question lie two different ways of conceiving the division of labour:
1. As the outcome of a single mind, which designs, divides and coordinates: this is the
view behind the Marxian capitalist and the Schumpeterian entrepreneur.
2. As the emergent, unplanned and unconscious outcome of the interaction of local and
incomplete decisions.
We believe that an intermediate form can be found in large business organizations, where
both forms, top-down conscious design and bottom-up partially unintended adaptation,
co-exist. To understand why this is so we have first to take into consideration the role
played by routines and contracts; both routines and contracts are necessarily incomplete
and partly tacit, and this implies that some discretion must be left to the actors. Thus
incompleteness and tacitness are likely to bring about cognitive conflicts.
To clarify this point, suppose that the top management puts into place a redesign of the
organization in order to react to some kind of environmental change. The implementation
of the new division of labour that is required by such a change gives rise to a complex
process of adaptation which is far from that anticipated by the traditional theory of
planning. On the one hand, the implementation of a new organizational design requires
managers and employees to rethink their jobs and revise their competencies, but on the
other handto be effectiveany new design requires local checks and adjustments, i.e.
the resolution of cognitive conflicts arising from a possible mismatch between the general
requirements of the project and the specific, idiosyncratic knowledge of any single agent.
These considerations lead us to a brief remark on the role of managers: for what has
been said, we cannot clearly distinguish between managing and doing:
the blue collar worker on the most routinized assembly line must
repeatedly make decisions about how to handle non-standard situations,
and in particular when to call one to the attention of the supervisor. On the
other hand, sales managers, in addition to managing their salespersons,
often spend considerable amounts of time with clients, engaged in selling,
and thus in doing.
(Radner 1992)
Doing, i.e. the application of a given routine, is never a mere execution of given rules
but always involves some discretionary behaviour, if only because a fully specified
routine, which does not leave any space for autonomous decision-making, would require
an endless list of contingent behaviours. The essence of managerial activities seems
rather to lie in the capability to fill knowledge and information gaps and in the capability
to design the organization of labour and imagine how to do.
Therefore top-down planswhich are necessarily incomplete for reasons we will
discuss belowhave to match pieces of knowledge and skills which are largely local and
tacit. Inevitably, they therefore have a largely uncertain and unintended outcome.
Before discussing the reasons for the incompleteness of planning, it is useful to note
that, when considered in connection with the process of division of labour, the very
notion of co-ordination takes on a meaning different from the one implicit in neo-
35
classical economics. For the latter, coordination means making individual and
independent decisions compatible; here the problem of co-ordination concerns the
relationship between the top-down activity of designing new organizational set-ups and
the adaptive, intelligent, bottom-up reactions by managers and employees, to adapt the
organization to the external environment.
Division of labour, distributed knowledge and specialization
The analysis of the process of division of labour as a primary determinant in economic
development goes back to Adam Smith and lies at the heart of classical economics. But
few of these early analytical efforts have been continued in modern neo-classical
economics, and they have been brought back to their central role only within recent nonneo-classical studies of the process of technological evolution (e.g. Dosi et al. 1988).
A possibly fruitful way of interpreting the process of division of labour is to consider
it within the general framework of the theory of problem-solving. From this point of
view, the division of labour derives from the decomposition of problems into
subproblems to be solved independently. Direct observation of the behaviour of
organizations and individuals as problem-solving activitywithin the tradition initiated
by March and Simon (1958)suggests that such behaviour is a peculiar and unstable
balance between two opposite situations: on the one side purely routinized behaviour, in
which a series of operations are repeated again and again, and on the other an active and
conscious search for solutions to new problems (or new solutions to old problems) faced
by the organization.
Without entering the debate on how to represent problem-solving activities formally,
some general points are worth mentioning for their relevance to the subject of this
chapter.
First, problem-solving activities are characterized by a search in a problem space
which generally leads to the decomposition of the original problems into at least partially
independent subproblems. If such a decomposition is feasible, subproblems can be solved
in parallel and subsequently co-ordinated: the original problem is therefore decomposed
into a set of connected subproblems. In the language of economics and organization
science, this corresponds to the expectation that the decomposition of a production
process (lato sensu) into independent subsystems, which can be dealt with in relative
isolation and in parallel, will lead to increasing returns, in spite of the co-ordination costs
which the decomposition generates.
Second, the working hypothesis we proposealong the lines of Simon and Marchis
that the search in the problem space, based on the division of the given problem into
subproblems, is a model for the division of knowledge; even if division of knowledge and
division of labour are not the same process (two employees co-operating in the same
organization can have the same competencies and largely overlapping knowledge bases
but different jobs), the former is necessary to the latter.
Third, it must be emphasized that the search in the space of sub-problems is a highly
uncertain and conjectural process. In fact, when a problem has been decomposed into a
set of subproblems, in general not all these subproblems will be solvable immediately and
some will have to be decomposed in turn into simpler ones. The decomposition then
Artificial societies
36
recursively proceeds until all the relevant subproblems have been solved. Problemsolving by subproblem decomposition is therefore an uncertain activity, for two reasons:
1. There is no set of decomposition rules which a priori allows agents to achieve a certain
result.
2. The solvability of the original problem can only be verified when all the relevant
subproblems have been solved (Egidi 1992).
This implies that plans and projects within firms are not merely executed by different
actors: they also necessarily require a multi-agent process of learning and adaptation that
can generate cognitive conflicts among the different actors.
These features allow us to consider co-ordination within organizations as the
complement to the division of knowledge and labour which follows the realization of a
new project. The further the division of labour proceeds, the more the different parts
require co-ordination and the more information becomes dispersed. A crucial problem
then arises: can we assume that the co-evolution of division of labour and co-ordination
also takes place within markets?
A first approach would claim that markets are mere co-ordinating devices (by means
of price mechanisms), whereas the division of labour is performed entirely within
organizations (Coase 1937). A more careful examination would suggest that the two
processes take place in a complementary way in both markets and business organizations,
the main difference being the degree of awareness which characterizes the actors.
The context of Schumpeterian competition is a good starting point to clarify how
division of labour and co-ordination can both take place in markets: suppose that a cluster
of product innovations is generated in the economic system. If the new products are
substitutes for old intermediate goods, they will activate the modification of the existing
productive and organizational routines and a second generation of adaptive innovators
will come along: they in turn, by using the new products will produce either new goods
or the old ones in a new manner. In the first case, we have a dynamic innovation process,
in which new projects lead to the modification of other projects. This is not merely a
matter of the diffusion of an innovation. On the contrary, what we are describing is an
avalanche of innovations that are activated recursively by the original cluster of
innovations. Two different responses could occur as a consequence of the original
innovative actions. In the first case, they only give rise to local dynamic processes, which
do not alter the basic structure of the division of labour. In the second case, the reaction
of the system causes a new definition of the division of labour, knowledge and
competencies within the economic system (often going beyond the intentions and
expectations of the innovators).
This clarifies how a new division of labour and knowledge can be gen-erated within a
market. It should be noted that co-ordination within markets also involves the
transmission of knowledge and competencies: prices do not convey sufficient information
to support the co-ordination processes that arise from the division of labour. The
importance of this point and the provisional conclusions we have reached are reinforced
if we consider an intermediate organizational set-up, as does Williamson in his
fundamental transformation (Williamson 1975). When two or more business
enterprises engage in a new common project, an improvement in human and physical
assets internal to each enterprise and a transfer of knowledge among the enterprises will
37
normally be required. The emergence of sunk costs in physical and human capital
guarantees the strength and stability of the links among the businesses. Even in this case,
the relationship among firms which takes place in the market will involve not only price
signals but also (and fundamentally!) the transmission of pieces of knowledge and
information. Conflicts and bargaining between the parties will most probably be solved
by protest and dispute instead of by leaving the market (Hirschman 1970). Markets are
only one of a number of devices for communication and co-ordination between
organizations.
To summarize, the transfer of knowledge and competence is a fundamental aspect of
co-ordination which takes place not only within organizations but also among
organizations in markets. Even if we believe that a fully developed model of the
properties of the division of labour and co-ordination and their co-evolution should be at
the top of the research agenda, to model this kind of process is far beyond our present
abilities and intentions. More modestly, we intend to pick up some of the properties
discussed above and try to compare the performances of the different co-ordination forms
that arise from the division of knowledge and labour in a context of boundedly rational
behaviour. In the next two sections the relevant assumptions will be discussed.
Division of labour and returns to scale
A fundamental question that has been already hinted at is whether and under what
conditions subproblem decompositionwhich we suggested was a useful representation
of the process of division of labourcan generate more efficient forms of the
organization of production. This question can be answered in many different ways: one of
the first analyses was provided by Simon (1962) with his parable of the two
watchmakers, called Tempus and Hora. Watches are complex objects made up of a very
large number of small pieces. Tempus assembles his watches sequentially and every time
he is disturbedfor instance, by his clients phone callshe has to start the assembly all
over again. Hora instead proceeds by first assembling subunits consisting of only a few
components and then putting the sub-units together. When he makes mistakes or is
disturbed by external events he only has to restart assembling the current subunit. His
average time to complete a watch will therefore be much shorter than Tempuss average.
In this example there is a clear advantage in the division of labour: when a
perturbation occurs in a sequential system, it affects the whole system, when instead it
occurs in a system made up of small independent and parallel subunits it will not
propagate outside the affected subunit. But if we rule out the possibility of perturbations
and mistakes, this advantage disappears and the parallel system seems rather less efficient
because of its higher co-ordination costs. Only if we consider the positive, long-term
effects that the division of labour might have on individuals working capability, because
of learning-by-doing and specialization, can we again find advantages in an increase in
the degree of task decomposition. Learning therefore seems to be a key factor in
explaining the division of labour. But how do individual learning processes co-ordinate in
a framework in which the very definition of tasks is an emergent property?1 In the rest of
this chapter we shall investigate this question by means of a very simple model of
division of labour and some simulations of different social co-ordination mechanisms.
Artificial societies
38
and put them in a new stack. To count this new stack, L2=100 accountants are
39
When the ratio log N/log k is not an integer, the number n of levels of the pyramid will
be the first integer greater than or equal to this ratio. This implies that the pyramid will
necessarily contain some slack resources which cannot fully be exploited.2 Let us
consider a general task of size N bills. If every accountant can count up to k bills, the
lowest level of the organization will produce N/k bills, the second N/k2 and so on up to
level w such that N/kk. If we have exactly N/K=k all resources in the pyramid will be
fully exploited, otherwise the productive capacity of the accountant at the top of the
hierarchy will not be entirely exploited. Moreover the numbers N/k, N/k2,, N/k might
not all be integers: some idle resources will appear at each level for which the ratio is not
an integer number.
The ratio between size of the task N and total amount M of labour required for its
completion is given by
R(k)=N/M=kn/1+k+k2++kn=(1k)kn/1Kn+1.
Thus R(k) tends to 1 as n tends to infinity and the process exhibits asymptotically
constant returns to scale.
If the number N of bills to be counted increases, the organization can respond by
increasing the level of employment and/or the productive capacity k. The latter can
increase also with N constant (equivalent to an increase of productivity, or Marxian
exploitation).
Artificial societies
40
The information requirements of this boundedly rational co-ordinator are quite different
from those of a fully rational central planner: while the latter needs precise information
about the size of the overall task and the characteristics of each accountant, the former
needs information on all the flows between different hierarchical levels. Whereas the
central planner needs precise information, as even small amounts of noise will make the
entire organization ineffective, the boundedly rational co-ordinator will generally need
some signals (even if they are only qualitative ones) about supply-demand disequilibria.
Moreover this boundedly rational co-ordinator can be replaced completely by inter-level
co-ordinators who each take care only of the relationship between demand and supply at
one interface between levels, regardless of what happens in other parts of the
organization.
As to the kind of cognitive capabilities that are required by the central planner and the
boundedly rational co-ordinator, the former has to develop a highly abstract and general
decision rule. This requires that the problem has been understood in its general features
and decomposed into subproblems. The boundedly rational co-ordinator, on the other
hand, can implement the organizational structure adaptively, using a process of trial and
error which can proceed without a general understanding of the problem (Dosi & Egidi
1991, Dosi et al. 1993).
But the process of adaptive co-ordination involves a cost, given by the loss of
efficiency incurred during the process of adaptation. While the perfectly rational central
planner computes the optimal organization in his mind, the boundedly rational coordinator carries out the design process in real time and corrects mistakes only after
experiencing the loss of efficiency they cause. On the other hand, if the central planner
makes mistakes, these are likely to damage the organization more persistently, because
the planner is unable to process signals which detect the presence of inefficiencies and
make consequential adjustments, and a thorough rede-sign is always needed even to cope
with small perturbations.
3. Co-ordination can also be achieved with a completely decentralized mechanism, a
quasi-market in which each accountant adjusts his production and/or his position
according to quantity and/or price signals, which are processed independently by each
individual. Each interface between levels of the organization constitutes a market
where the accountants of the lower level sell the fictitious bills they produce and the
accountants of the higher level buy them. Demand and supply in each of these markets
depend on the number and productive capacity of the accountants at the two levels.
Suppose, for instance, that the overall productive capacity at the lower level is not
sufficient to supply the accountants at the higher level with enough bills. Some of the
higher-level accountants will not be able to produce enough because of insufficient
input and will therefore tend to move to other parts of the organization. These
adjustments could take place through simple quantity signals or through more complex
price mechanisms of Marshallian type (for instance, excess demand generates a price
increase which raises the profits of sellers; and this attracts new sellers who make the
supply increase and balance the initial excess demand).
A decentralized co-ordination mechanism of this kind requires only local information
processing: each accountant can process disequilibrium signals autonomously according
to his own local knowledge about his own characteristics and without any need to know
41
Artificial societies
42
levels of the hierarchy. It can thus be represented by a set of condition-action rules where
the condition classifies such signals and the action defines a complete organizational
structure. In detail, we have:
(a) Environmental messages (equilibrium/disequilibrium signals) are given by the
concatenation of eight binary strings (one for each interlevel interface, including the
interface between the last level and the final demand). Each one of these eight
substrings is composed of two digits:
s1s2 s {0, 1}
where the first digit is set to 1 when there is an excess supply at that interface and
is set to 0 otherwise; the second digit is set to 1 when there is an excess demand at
that interface and is set to 0 otherwise.
(b) Conditions are strings of the same length (16 bits) as the environmental messages
which they classify:
c11c12c21c22c81c82 c {0, 1, #}
Each bit position may be set to 0, 1, or a wild card marker, #.
(c) Action parts are binary strings of length 6h that define the whole organizational
structure:
p11p12p13k11k12k13ph1ph2ph3kh1kh2kh3 p, k {0, 1}
In this way an adaptive and boundedly rational central co-ordination can be represented
by a classifier system. Here we briefly review its basic features (more details can be
found in, for instance, Holland 1986, Holland et al. 1986, Goldberg 1989).
An adaptive learning system is a system of condition-action rules such as those so far
described. In addition, each rule is attributed a strength coefficient which, as a first
approximation, measures how successful that rule has been in the past, and a specificity
coefficient (the number of bits in the condition part that are not set to the wild card #),
which measures the strictness of the condition. The smaller the cardinality of the set of
environmental messages that satisfy that condition, the higher is its specificity, the
highest specificity belonging to rules whose conditions are satisfied by only one
environmental message.
This set of rules is processed according to the following cycle throughout the
simulation process:
1. Condition matching: a message is received from the environment which informs the
system about the disequilibrium/equilibrium conditions at the inter-level interface.
This message is compared with the conditions of all the rules, and those rules which
match, i.e. those which apply to such a state of the world, enter the following step.
2. Competition among matched rules: all the rules whose condition is satisfied compete in
order to select the one which is allowed to execute its action, i.e. to implement the
organizational structure specified by its action part. To enter this competition each rule
43
makes a bid based on its strength and on its specificity. In other words, the bid of each
matched rule is proportional to its past usefulness (strength) and its relevance to the
present situation (specificity):
s1s2 s {0, 1}
Artificial societies
44
where the first digit is set to 1 when the accountant has been rationed on the
demand side (i.e. there is an excess demand at the interface between that
accountants level and the lower one) and is set to 0 otherwise; the second digit is
set to 1 when the accountant has been rationed on the supply side (i.e. there is an
excess supply at the interface between that accountants level and the higher one)
and is set to 0 otherwise.
(b) Conditions are therefore strings of two bits which classify such environmental
messages:
c1c2 c {0, 1, #}
(c) Action parts are binary strings, each six bits in length which define the accountants
position in the hierarchy and his productive capacity:
p1p2p3k1k2k3 p, k {0, 1}
Each classifier is then processed by exactly the same execution cycle as that
already described for the centralized co-ordination case.
The two co-ordination systems have been tested on a simple problem. The task is to count
25 bills and there are 6 accountants, whose capacity can vary between 0 and 7 banknotes
and whose position in the hierarchy can vary from level 0 to level 7. Level 0 is a stand-by
position: accountants in this position do not enter the production process5. Let us examine
in more detail the nature of the environmental signals which characterize the two
institutional set-ups:
1. In the case of the boundedly rational central co-ordinator, the winning rules action
part is decoded in order to obtain the corresponding organizational design. The
productive capacity of all the accountants who are at hierarchical level 1 is then
summed to obtain the total demand for banknotes at this level. If the total demand is
less than 25, the environmental message will signal an excess supply in the follow-ing
iteration (the first digit will be set to 1 and the second to 0). If the total demand is
greater than 25, the environmental message will signal an excess demand (the first
digit will be set to 0 and the second to 1). Only if the total demand is equal to 25 will
the environmental message signal an equilibrium situation (both digits set to 0). The
total supply of level 1 can now be computed in this way:
(i) if the total demand is less than or equal to 25, all accountants at level 1 will be able
to exploit their productive capacity to the full. Thus the total supply will be given
by a number of banknotes equal to the number of accountants, and each note will
have a face-value equal to the productive capacity of the accountant who produced
it; and
(ii) if instead the total demand is greater than 25, some accountants (randomly chosen)
will be unable to use their own productive capacity fully. Total supply will be given
by a number of banknotes equal to the number of accountants at the first level who
received at least one banknote and their face-values will be given by the production
of the accountants who produced them.
45
Once the total supply at the interface between levels 1 and 2 has been so computed, we
can determine the total demand as the sum of the productive capacities of all the
accountants who are placed at level 2 of the hierarchy. We can then set the third and
fourth digits of the environmental message according to the disequilibrium and
equilibrium situations which are thus realized. The same procedure can be repeated for all
the organizational levels. At the last level we will suppose the existence of a final
demand of one banknote of face value 25. If more banknotes are offered, an excess
supply will be signalled and the face-value of the only purchased banknote will determine
the overall payoff to the organization: if its value is 25, the organization (i.e. the winning
rule) will receive a positive payoff, otherwise it will receive a negative payoff
proportional to the absolute difference between 25 and the face-value itself.
2. In the case of decentralized co-ordination, inter-level supplies and demands are
computed in exactly the same way, but environmental messages are determined
separately for each accountant, by means of random rationing. For instance, if at a
given interface between levels, demand is higher than supply, banknotes are assigned,
one by one, to a randomly chosen accountant who still has some unused productive
capacity. All accountants on the supply side will therefore have sold all the banknotes
they produced, and at the following iteration will receive a message whose second
digit is set to 0. As to the accountants on the demand side, some (possibly none) of
them will have been given all the banknotes they required and at the following
iteration will receive a message whose first digit is set to 0. The others (possibly all of
them) will find themselves with at least a part of their demand unmet and at the
following iteration will receive a message whose second digit is set to 1. The
organizational payoff is in this case distributed to all the accountants winning rules
through a bucket-brigade mechanism (Holland 1986).
Simulations have been carried out in order to test the adaptive performance of the
systems in different environmental conditions. Let us ex- amine briefly the main results:
1. The first set of results concerns the simplest situation: an error-free and stationary
world. The stack of banknotes to be counted always contains 25 notes, and no
mistakes can occur in the counting process, i.e. the counting can always be performed
without any disturbance and the face-value of the output always equals the amount of
used productive capacity. In this situation the boundedly rational centralized coordination mechanism is considerably more efficient in finding the optimal structure
as far as both the speed of convergence to an effective organizational design and its
stability are concerned.
The structure of the payoff function was found to play a crucial role in
determining the speed of convergence and the type of organization which
emerges. In particular, punishments for unused productive capacities are essential
for reducing slack resources, and a payoff function that takes into account the
total task-completion time is necessary to obtain an equal distribution of
capacities across agents.
2. In error-prone environments, accountants can make mistakes. Two different kinds of
mistake are possible. In the first case accountants are aware of the mistakes they make
and correct them by restarting the counting process. The payoff function contains a
Artificial societies
46
penalty for the time (steps of the counting process) taken to perform the overall task.
A second type of mistake is represented by a random variable (with mean 0) which
might cause a deviation between the amount of productive capacity used by an
accountant and the face-value of the banknote he produces. In this case accountants
are not aware of the mistakes they make and they deliver a result which is not correct.
The decentralized mechanism is more efficient in dealing with both types of mistake.
Local adjustments therefore seem to increase efficiency in coping with local
disturbances.
47
Notes
1. The problem of the co-ordination of learning processes within an economic organization has
also been studied, in a framework in which the division of labour is given, in Marengo
(1992).
2. The presence of such idle resources is a possible reason for the existence of economies of
scale in this technology: a multiplicity of tasks can be internalized in a single organization,
which can therefore reduce the amount of idle resources. In the above example, two separate
would be handled more efficiently by a single organization rather than
tasks of
separately. But this issue lies outside the scope of this chapter.
3. This requirement seems rather paradoxical. If the planner knew exactly the number of bills to
be counted there would be no need to carry out the counting process. This paradox also
arises in less caricatured production processes, although less strikingly. Fully centralized and
exhaustive planning would require perfect knowledge of every aspect of the production
process: only in this case could the co-ordination problem be solved in the planners mind
without the need for local adjustments.
4. Markets with prices are much more complex and to model them correctly we should
introduce a series of hypotheses on individual utility functions, which would take us far from
the core issue of the chapter.
5. The possibility of leaving the organization must be allowed for if we want the system to be
able to adjust when there are more accountants than the optimum number.
Chapter 4
Emergence of kinship structures: a multiagent approach
Jean Pierre Treuil
Among the various systems of representation that form a culture, those concerning
kinship and spatial relationships are obviously highly significant. These have been prime
elements in human exchange and they are closely related in many societies. This chapter
takes as its starting point the kinship systems of the Australian aboriginal societies
studied by B.Glowczewski, an anthropologist. It reports the first part of more extensive
research conducted jointly with him. The research aims at studying, by means of
computer simulations, the conditions governing the observed structuring of territory in
these societies.
49
and marriage, starting from an original society composed of individuals with differing
rules. In order to formulate such a process, the interactions between systems of rules must
be capable of description. Furthermore, the systems of rules must belong to the same
general pattern, each representing a specific development of it.
The process described here leads to systems of kinship rules which share a significant
property with the elementary kinship structures described in the literature, that of being
mathematically formalizable in terms of permutation groups. Our kinship structures
also share another important property with the real ones: the property of exogamy.
However, the real systems also exhibit properties of symmetry (see the example of
Kariera society below); these properties ensure that all the matrimonial classes identified
in a given kinship structure have an equivalent position and that no one has a privileged
status in the system. Such properties are more difficult to generate in our simulations.
Nevertheless, we think that the simple hypothesis at the basis of convergence, namely an
evolution which searches permanently for a consensus between the individuals who make
similar choices, can provide some interesting explanations of these cultural phenomena.
Artificial societies
50
l3
l2
l4
24
21
23
31
34
32
42
43
41
51
Artificial societies
52
53
Conditions of union
The conditions of union, , are based on a principle of agreement between the
individuals selected from the potential couples about the classes of their children. This is
a generalization of the relationship =
The simplest principle only requires that the maternal part of the model of the mother
and the paternal part of the father should be consistent, at least with respect to the class of
their children. Let i and j be the potential parents. can be written
Ri[si,xi]Rj[sj,xj]0
Then
will designate the possible children of the couple.
A more limiting variant of this condition requires first the separate agreement of the
fathers and the mothers models. Thus, under this condition, a man of class xi should
marry a woman of class xj. This condition can be written:
Ri[si, xi]Ri[sj, xj]Rj[si, xi]Rj[sj, xj]#0
Meeting protocol U
We experimented with a protocol which runs as follows:
1. Individuals first choose a filiation for their future child. This choice is made from all
the filiations allowed by their own model.
2. Couples are selected to breed at random from among those who made similar choices.
The procedure we used, called the veto procedure, requires that the choice made by an
individual i from a filiation z must not be prohibited by another individual of the same
sex who has already chosen this filiation z. In more detail, the process is as follows:
An individual i first chooses a filiation z at random from his or her own set Zi=Ri[si,
xi]; then i asks individuals of the same sex who have already chosen the filiation z if they
agree with this choice; if one of these individuals j does not have the rule (si, xi, z) in his
or her own model Rj, the individual i will search for another filiation in Zi. Moreover, in
what we call below Lamarckian processes, it will remove the rule (si, xi, z) from its
model Ri, so that it will never again choose the class z.
This kind of procedure is sophisticated but it corresponds to a situation which is not
unknown in human behaviour. Its application has an important consequence: it reduces
the diversity (in terms of classes) of the set of individuals choosing a given filiation. This
set becomes increasingly specific with time.
Combination operator C
The principle consists in establishing the maternal and paternal relations of the child,
based on the parents maternal and paternal relations, respectively.
Let i be the father and j be the mother of the child, k:
Rk=C[Ri, Rj] Pk=C[Mi, Mj] and Mk=C[Pi, Pj]
Artificial societies
54
55
belonged. The length of the cycle is the size of this set. These cycles are characteristic of
the average structure of kinship models over the population.
Artificial societies
56
Results
All the experiments analyzed here are carried out with eight classes and a population of
256 individuals. The choice of eight classes is consistent with many observed structures
of kinship. We have also carried out experiments with 16 classes and more numerous
populations, but the results are not essentially different in their qualitative features. We
discuss later the influence of the number of individuals.
Convergence
We observed the evolution of the above measures for the same periods for each kind of
simulation. We analyzed the results from the values of:
57
1. The mean number of possible filiations in the individuals models, F(t). The maximum
value of F is 288=128 filiations, and the minimum is 0.
2. The mean number of possible classes of the children of the actual unions is C(t). The
maximum value of C is 8. The minimum is 1, otherwise a union would be not realized.
3. The entropy of the distribution of the population on the set of classes E(t). Its range is
between 0 and 3.
4. The average specificity of classes S(t). Its range is also between 0 and 3.
Within any one kind of simulation, we observed that F(t), C(t), E(t) and S(t) always have
similar shapes. These variations are related to discontinuities within the evolution of
classes and the structure of R. The curves and observation of the process suggest that it is
useful to break the simulation into three phases (see Figure 4.3).
1. During the initial phase, F and C decrease evenly at a speed linked to the mutation rate.
They are related by a formula which comes from the hypergeometric law governing
the statistical distribution of C when F is given: C=F2/4N3. During this phase, there
are meetings between all classes; there are many non-zero values within the union
matrix. So the sizes of all the classes change very slowly. Entropy remains close to its
maximum value, 3, for all eight classes. This phase ends when C reaches a value close
to its minimum, i.e. close to 1. At this point the individuals who meet to reproduce
usually get only a single common possible filiation: the one they choose.
2. During the second phase, corrections and stabilization can be observed: C remains
constant, slightly over its minimum value; F and S continue to decrease until they
reach minimum values which are related to the mutation rate; E remains close to its
maximum or falls into a lower value. Many null values appear within the union
matrix. Individual kinship models become similar, but their common structure does
not correspond to a permutation group because different classes may have the same
filiation. The union matrix is not stable. There are
58
Artificial societies
59
It is not always easy to determine the end of this phase precisely. S(t) is the
relevant indicator, but with high mutation rate, noise is important. The earliest
point at which the phase ends is when S(t) stabilizes.
3. The third phase is the equilibrium phase: C has reached its minimum value and
remains constant. All the individual kinship models are approximately the same. This
common model Rm corresponds to a permutation group (a bijective mapping of the set
of occupied classes). The cycle structure is clear and remains constant. However,
observation shows that even permutation groups are not completely stable. This
residual unstability occurs in Lamarckian processes with high mutation rate, when the
permutations do not cover all the initial classes.
We can observe three variations of these common features. They are all related to the
second, stabilization phase:
Variation 1. Stabilization ends suddenly with a dramatic fall in the number of occupied
classes (see Figure 4.4(a)). It sometimes results in the termination of the reproduction
process.
Variation 2. The decline in entropy ceases and future adjustments result in the even
distribution of the population over the classes (Figure 4.4(b)).
Variation 3. The situation does not stabilize. (Figure 4.5).
Table 4.1 summarizes the results of the 40 simulations.
States of the system after stabilization
The states of the systems after stabilization are shown Table 4.2. This table gives the
equilibrium values of F (average number of filiations per class in the individuals
models), N* (the number of occupied classes), and the number and lengths of the
stabilized cycles for women and men. For example, the third simulation with a
Lamarckian principle, a low mutation rate and limitation of the class size is shown in
Figure 4.3(a).
For this simulation, in Table 4.2, we have F=18, N*=7=N1 (one unique class is
removed). There are two cycles (lengths 2 and 5) for women and one cycle (length 7) for
men.
Synthesis
Tables 4.1 and 4.2 allow us to provide a quantitative answer to several questions.
Artificial societies
60
61
Artificial societies
62
7000
7000
7000
8400
9700
9700
9700
9700
9700
9700
9700
9700
9700
9700
700
600
700
700
700
350
500
350
350
350
1800
1800
2500
2100
1400
1000
700
4800
4800
4800
4800
4800
4800
4800
4800
4800
4800
DARWINIAN
2100
LAMARCKIAN
DARWINIAN
End of the init. phase
900 1400
1400
8400
8400
4000
2100
2800
1400
5300
9700
9700
9700
9700
9700
9800
9800
9800
9800
9800
700
600
700 1000
1000
350
350
350
200
350
1800
2100
2600
2800
1800
1000
1000
4800
4800
4800
4800
4800
4800
4800
4800
4800
4800
LAMARCKIAN
63
DARWINIAN
Number of filiations F
21
25
26
Remaining classes N*
Womens cycles
1+6 1+2+3
Mens cycles
2+5 2+2+2
Incestuous filiations
29 30
42
42
44
40
40
1+1+2
26 23
20
23
28
33
40
Number of filiations F
17
18
18
Remaining classes N*
1+2+5
2+6
2+5
1+4
Mens cycles
4+4
Incestuous filiations
LAMARCKIAN
Womens cycles
7 1+1+3
1+2+5
DARWINIAN
Number of fliations F
32
36
22
27
50
56
57
61
41
Remaining classes N*
Womens cycles
3+3
Mens cycles
2+4
1+1
Incestuous filiations
LAMARCKIAN
Artificial societies
Number of filiations F
18
18
17
Remaining classes N*
Womens cycles
1+7 2+2+4
Mens cycles
3+5
Incestuous filiations
64
24
27
37
55
47
1+5
1+7
1+6
8 1+3+4 2+2+2
1+7
2+5
1+4
20 17
3 1+2
0
For high mutation rates, the same observation can be made. In stabilized situations where
all the initial N classes remain occupied, the level of F is also very low. The average level
of filiations becomes relatively high in situations where many classes are unoccupied,
and therefore where many inactive rules exist. But even in these cases, the number of
possible filiations in active parts of the model remains near the minimum.
Does the number of occupied classes remain constant?
The procedure of veto involved in the chosen protocol of meetings tends to eliminate
rules from the individuals models, in both Darwinian processes and Lamarckian ones.
Does this removal of rules keep the number of occupied classes near the initial number?
Here again, a general feature can be observed. Evolution tends to keep a high number of
occupied classes, as may be observed in Table 4.3(a) which gives the relationship
between the stabilized situations and the number of occupied classes, N*. In a majority of
cases, the number is greater than 5, even without any constraint on its minimum value.
Simulations without
restriction on the class size
65
10
13
10
66
Artificial societies
67
offset the disturbing mutations. There is a sort of learning which leads to preserve a larger
number of occupied classes.
Discussion
The questions that have been raised in this chapter are as follows:
1. Is it possible to build a process based on interactions between individuals chosen
stochastically that gradually homogenizes their individual kinship models by making
them consistent with a common model?
Artificial societies
68
2. Is it possible to choose conditions for the interactions so that this common model is
consistent with models of elementary kinship structures?
3. Is it possible to achieve this under conditions based on what is known about the
history of such structures?
Simulations allowed us to answer the first question affirmatively. They also showed
that a certain type of protocol which does not seem to be unrealistic leads to models
which prescribe exactly the unions and filiations consistent with the preservation of
matrimonial classes and the constant renewal of the population. On the other hand, the
symmetry properties of elementary kinship structures which allow an equilibrium to be
reached between the different groups and which retains their status in successive
generations could not be obtained. Moreover, it is not certain that they could be obtained
naturally by remaining in the same schema, that is to say, without making use of
additional knowledge in individuals models. Finally, the negative answer given to the
third question about timescales leads us to think that the models diffusion and
homogenization follow courses other than those used in these simulations.
However, we think that the simulations are of interest. They show that structuration
can be established from very simple conditions. The factor common to these conditions is
a transfer of information. The individuals not only act according to direct information
affecting their behaviour, but they also take account of the rules of the others with whom
they become acquainted during meetings.
Our approach consisted of first making a generalization on the basis of observed
behaviour and then testing the processes and the modes of transmission and change that
lead from an original society displaying all possible behaviours to the eventual
prevalence of that observed behaviour. We shall now introduce a theoretical reflection
about this way of explaining cultural evolution.
69
are none the less regarded as the consequences of hypothesized mechanisms of individual
behaviour.
The holistic approach accentuates the influence of the whole society on its
constituent parts and the way society determines an individuals degree of autonomy. The
determinants that are brought into play vary with the changing trends of holistic thought.
Some have tried to explain the rules of society by proposing that the rules enable the
society to reproduce and survive by adapting to the changing environment in which it
exists. More generally, rules can be considered as the ways used by the society to fulfil a
certain general condition. For example, in family systems, marriage rules may be linked
to the need to prevent consanguine groups from becoming introverted and to maintain a
balance between what each group gives and gets (Levi Strauss, quoted in Mercier 1968).
Caill (1992) argues that the individualism/holism debate (see also Chapter 8) should
give way to the idea of collective subject because a society builds itself through
choosing its environment and therefore the constraints within which it must live; through
choosing the values which it will nurture, in view of these constraints; and through
choosing the solutions adopted to resolve the tensions created by these constraints and
values.
Tautology in simulations
The charge of tautology has been brought against both the individualistic and the holistic
approach (Caill 1992). In the former case, the tautology can be found in the link
between the choice of the individuals and the goals they pursue. In the holistic approach
it arises from the link between the problem to be solved and the means used to solve it.
Similar criticisms are made of computer-simulated social processes: if a machine
generates an artificial world, what is observed can only be the result of programming the
rules which govern its existence.
Some methodological precautions must be taken to test the tautological nature of an
explanation; they are essentially the same whether computer simulation is used or not.
But experiments that can be carried out with the use of a computer do give the researcher
an excellent chance to test the results.
In our opinion, the following methodological steps need to be applied in carrying out
research by simulation:
1. Identify in the social processes general properties: satisfied constraints, objectives
pursued or values to be nurtured.
2. Characterize these general properties by measurement or qualitative means so that their
fulfilment can be checked in a real or simulated society.
3. Delimit a space of possible and concurrent behaviours.
4. Determine within this space those types of behaviour that are incompatible with the
general properties identified, as well as those that encourage the emergence of these
properties, and analyze the minimum hypotheses required for this emergence.
Artificial societies
70
71
elements. It is difficult to see how a partial and random modification of such a structure
could maintain its consistency.
At this stage, it seems useful to put forward the following hypothesis: human
knowledge and representations are arranged recursively in a system of hierarchically
organized schemata, each of them delimiting a set of possible changes. Cultural evolution
would be analyzed as the result of a succession of permitted changes within schemata, at
a different level of this organization. Some of the changes, occurring initially in a few
individuals, could be transmitted to others through the usual modes of transmission and
under some conditions, snowball to spread to the whole society.
Chapter 5
Cities can be agents too: a model for the
evolution of settlement systems
Stphane Bura, France Gurin-Pace, Hlne Mathian, Denise Pumain
and Lena Sanders
Models of global change sometimes take into account the population dimension, with
various conjectures about demographic growth. More rarely they include hypotheses
about the spatial distribution of the population, that is, the geography of settlement
patterns. This neglect is regrettable because most of the relationships between societies
and their environment depend on the nature and size of human settlements, at the microlevel of the individual as well as at the macro-level of regional planning. In particular, the
general and powerful contemporary trend towards an increasing concentration of
population within city systems in many countries of the world has to be considered
seriously (Moriconi 1993). The grounds for such a process are not yet fully understood:
for instance, many authors have interpreted the negative correlation between
demographic growth and settlement size in the industrialized countries during the 1970s
as initiating a new trend, called counter-urbanization (Berry 1976). The subject is still
controversial (Champion 1989).
Settlement patterns have been a source of concern for geographers for a long time.
However, a conceptual link is still lacking between the comprehensive theories of
urbanization describing the transformation of a rural population into an urban one, and
the genesis of the corresponding settlement patterns. It is not clear how a highly skewed
size distribution and heavily concentrated spatial pattern of towns and cities emerges
from a more homogeneous and more scattered set of rural settlements. Comparative
analysis suggests that developing countries may follow a process quite similar to the one
that was observed in the industrialized countries during the past two centuries, although
involving different parameters values. A model simulating such a process would be
useful to test hypotheses about the key parameters of the dynamics of settlement systems,
and to predict the possible evolution of systems in various demographic, economic and
environmental conditions.
Central place theory is the major conceptual framework that has been employed in this
field but, although its dynamic aspects were mentioned in early work (Christaller
1933:86128), more recently they have received only intermittent attention from
modellers. Central place theory deals with the regularities found in the spacing of
settlements according to their rank in a hierarchy of size. The theory suggests that the
cities with the most important functions (e.g. administration) are spaced further apart than
cities with only simple functions. Recently, several authors have proposed to simulate the
development of central place systems by means of mathematical dynamic models based
upon differences (White 1977, 1978) or differential equations (Allen & Sanglier 1979,
73
Gamagni et al. 1986). These models use a limited number of equations, defined for each
state variable and the same for every spatial unit. They are based on general hypotheses,
defined at the macro-geographical level, about the way urban change is occurring in the
system.
In Whites model, the relationships between retail activities, described by cost
equations, and consumers are simulated by spatial interaction equations. He links the
spatial dispersion of centres to an interaction parameter and tests the compatibility of the
urban hierarchy with a rank-size distribution for the cases of one-sector (White 1977) and
two-sector (White 1978) simulations. Allen and Sanglier (1979) use differential equations
for simulating the more or less random appearance of central functions at higher and
higher levels and their competition within a system of settlements. However, although
these authors get final distributions of city sizes which satisfy the rank-size rule, the
temporal developments of the spatial patterns of centres are more typical of a market
progressively colonized by entrepreneurs rather than the genesis of a realistic central
place system, since the models start with two central locations only, sharing the whole
region as their market areas (Allen 1978).
Such models seem well adapted for simulating competition between already
established urban centres. Modelling with differential equations has some advantages: the
model is written in a rather concise way and it can be experimented with under specified
conditions with a guarantee of repeatability. However, such models treat geographical
space as an isotropic function of distance, using various synthetic spatial interaction
formulations; they do not allow the mixing of quantitative and qualitative information;
they have difficulty handling a large variety in both the range and scope of spatial
interactions; and they cannot consider more than two different geographical scales at the
same time. It would have been difficult to adapt any of these models to our project in
which we consider the historical development of a set of cities over a long period of time,
involving several qualitatively different functional types of city.
That is why other kinds of modelling have been tried in studying the evolution of large
settlement systems. Among the first attempts was the pioneer work of Morrill (1962),
using a type of Monte Carlo simulation for modelling the development of a central place
system, a method first used by Hgerstrand (1952) for simulating the spatial diffusion of
innovations. Progress in computer science and artificial intelligence (AI) has now opened
new avenues to spatial modelling: for example, with cellular automata. A cellular
automaton consists of an array of cells, each of which may be in any one of several states.
At each iteration, each cell may change to another state depending on the state of
neighbouring cells. Some models, like the famous Game of life (Berlekampe et al.
1982), are very simple, but more complex cellular automata could be useful in
geography, if they involved more than two possible states and if they allowed for more
sophisticated definitions of neighbourhood. Tobler (1979) first mentioned them as the
geographical type of model, while Couclelis (1985) drew attention to their use for
modelling micro-macro relationships within spatial dynamics and for deriving complex
dynamics from simple rules (Couclelis 1988). White (1991) applied this formalism to
simulating the evolution of land use patterns within urban areas.
Our research relies on the same basic principles as sophisticated cellular automata but
has benefited from an advanced modelling method that is known in the field of AI as
multi-agent modelling (Ferber 1989). Compared to cellular automata, multi-agent
Artificial societies
74
modelling more easily takes into account a large diversity of interactions between the set
of cells or objects under consideration. In particular, it can include definitions of
neighbour-hoods that are more flexible and more suitable for the study of urban systems.
In this chapter, multi-agent models will be described first, and we shall see how they
can be applied to geographical simulations. A theoretical framework for the development
of settlement systems will then be presented, so that objects and rules can be chosen for
the design of the model.
75
units of the system can be modelled. The differences of range that depend on the level
of the city in the urban hierarchy can easily be taken into account;
(c) the behaviour of an agent is the result of its observations, its knowledge and its
interactions with the other agents: the rules governing the evolution of each agent
are also handled locally. These rules make it possible for each unit to define its own
evolution from knowledge of its actual state and of its exchanges with other agents.
This formalism makes it possible to consider the city as an evolutionary system. It
receives and diffuses information; and it can create and innovate. It can also transform
its qualitative functional structure: for instance, a hamlet or a village can become a city
at an increasingly higher level in the urban hierarchy. As it competes with others, the
reverse can also occur; and
(d) an agent is a real or abstract entity that is able to act on itself and its environment.
An aggregate of population acts on its environment of course; for example, by
exploiting its resources, or building transport infrastructure. The effectiveness of its
action will depend upon the size of its population and of its functions.
A multi-agent approach differs in many respects from the sets of differential equations
that are often used in dynamic modelling. The differences appear both in the treatment of
the data and in the working of the model. An example helps in understanding these
differences. The negative effect of distance is a frequently-used component in spatial
interaction models. In a differential equation approach, this effect will appear as a
function of distance (for example, a negative exponential) inside a global equation that
combines many effects and which describes the evolution of various state variables. In a
multi-agent approach, the level of generality will be the same: the rules are defined in the
same way for all agents of the same kind, but the evolution of the systems state is
handled at the level of each geographical unit. In a multi-agent framework, each
geographical unit is able to identify the other units of the system, their proximity and
their properties by exchanging information. The volumes of migratory flows or trade
exchanges between two cities will be computed at the scale of the city, using the rules
that are associated with its status as a city, for example. Although the rules have a quite
general definition, they are used and handled at the local level of the geographical units
that are the building blocks of the model.
The most useful quality of multi-agent system modelling is its flexibility. It is easy to
introduce modifications: it is possible, for example, to introduce new behaviour. This
kind of change involves the construction of new rules but does not entail changes in those
that already exist. Each enrichment of the model is thus made by adding rules that do not
affect the global structure of the system. It is even possible to create a new class of agents
(a transport network, for example) which is able to communicate with the city agents. In
particular, it is possible to test the effects of individual decisions or specific local
strategies without modifying the general rules. One could, for example, test the effect of
the past-orientated attitude of a town council that hinders a town from growing as quickly
as its potentialities might allow. The handling of special cases is very simple as each
agent has local knowledge of the general rules that are associated with its status and the
specific rules that correspond to its particular case. In this way, it is possible to introduce
simultaneously into the model some rules that depend on general laws (e.g. distance
decay, competition, supply and demand mechanisms) and others that reflect local
peculiarities.
Artificial societies
76
With this kind of modelling it is also easy to handle geographical units that are of
various kinds and can change from one qualitative state to another (e.g. hamlet, village,
cities of various levels). In this way the genesis of a city and of an urban system can be
modelled.
In many respects, this way of conceiving of the evolution of a complex system seems
relevant to modelling of the evolution of the urban system. Our purpose is not to study
the development of one city but to analyze the evolution of the structure of the whole
settlement system. In order to transfer this methodology to our problem, we have to recall
the main evolutionary properties of an urban system, the purpose of the next section. The
model is supposed to reproduce such an evolution from the modelling of the interactions
at a meso-level. The micro-level of individual actors will not be considered here.
77
only the biggest cities accumulated the functions and the hierarchy got simpler
because of the growing differences between the cities and smaller settlements (GurinPace 1993).
The historical genesis of urban systems from a dispersed pattern of rural settlement may
be interpreted in terms of two alternative theoretical explanations. The first claims that
towns emerged endogenously as villages accumulated wealth and developed central
functions. Variants of such a bottom-up theory either consider that this process is
economic by nature and that the emerging towns were at first rural markets exchanging
surplus of agriculture within small regions (Bairoch 1985); or that the selection of
specific nodes was political, as feudal lords decided to group in particular places all the
craftsmen they needed for their own convenience (Duby 1984). According to the second
theory, the emergence of towns and cities was linked from the very beginning with the
development of long-distance exchanges. Large cities could grow without being
supported locally by a rich agricultural production, because they had an especially good
situation for long-distance trade (Pirenne 1925, Braudel 1967). The very general process
of hierarchical diffusion of innovations within urban systems could support this second
hypothesis. Both explanations, bottom-up and top-down, can be exemplified in the
literature about urban history and can be tested in our model.
The main objective of our work is, however, to test the consistency of our hypothesis
about the self-organized character of settlement systems: as open systems, the dynamics
of their structure at the macro-level are supposed to be generated by interactions of the
elements of the system at a meso-geographical level. By simulating the interactions
between the local units of a set of settlements, we should be able to produce an overall
evolution of the whole system that is compatible with the state of our knowledge about
settlement systems dynamics, as enumerated above.
Despite the irreducible uniqueness of the historical path of development of each
human settlement (Arthur 1988), it is possible to identify a few rules that may
characterize the successive transitions leading from a small agricultural settlement to a
large multi-functional urban metropolis. In addition, other rules describing the way
settlements interact during their development are needed.
The main transition within the history of a settlement occurs between villages and
towns, which are quite different as geographical entities. Villages exploit local renewable
resources, levied on their own site. Their ability to sustain and to develop their wealth
and their population depends mainly on the quality of the available resources; they are, of
course, affected by such external events as climatic fluctuations and military invasions,
but they have very little power over such perturbations. A kind of Malthusian equilibrium
limits the growth of population according to agricultural production. However, villages
may accumulate the surplus of production and commercialize it.
Some of those settlements may then become towns by developing a totally different
way of accumulating wealth and sustaining the population. Trading ability or political
power is used to capture the wealth produced on distant sites, by means of unequal
exchange. Production is also diversified by inventing new products and new skills. The
ability to accumulate wealth no longer relies on their own local resources but on the
quality of the settlements situation within communication networks. The constraints on
development depend mainly on competition with other cities for markets and levying
taxes on other sites. Town markets are at first confined by transportation costs to a small
Artificial societies
78
neighbouring area, but they enlarge over time with progress in transportation facilities.
The scope of market trading also enlarges as towns grow and create more activities that
are new urban functions. The neighbourhood is then defined in terms of topological
and hierarchical distances, measured on the transportation and information networks
between urban centres, instead of consisting of a restricted contiguous market area. To
our knowledge, this is the first time that the effect of space-time contraction has been
introduced explicitly within a model of central place dynamics.
A complete model of settlement dynamics should therefore include explicitly at least
three kinds of interaction flow between places: wealth, population and information. In our
model, each place, according to its natural resources, its functions and its position within
the system, produces wealth. Differentials of wealth between places and the circulation of
wealth through exchanges induce the development of the settlement system. The effect of
wealth accumulation on population growth is formulated in a simplified way: the ability
of a place to increase its population depends on its own wealth. In order to avoid the
deterministic effects of that too-simple assumption, population growth rates are computed
as stochastic functions of wealth (in the first version of the model, population migrations
are not considered explicitly but could be integrated later). The circulation of information
is another important component in the dynamic of the multi-agent system: information
about available production is provided to neighbouring consumers, and information about
existing demand to trading centres. During the course of development, innovations and
new economic functions are created, through which places gain the means of producing a
wider range of goods with a larger value added. Such information circulates between
towns and cities according to a hierarchical definition of their neighbourhood. Therefore,
the circulation of information between places allows the circulation of wealth. It also
encourages competition for accumulating and producing wealth.
In a further step, population flows could be added to the model, following, for
example, the suggestions made by Huff (1976) or by Fik (1988) and Fik and Mulligan
(1990) about the distribution of migration flows within a central place system. A more
general synergetic model of inter-urban migrations could also be included (Haag et al.
1991, Frankhauser 1990, Sanders 1992).
The model
This application has been developed in the object-oriented language, Smalltalk, because
of its modularity and flexibility (Goldberg & Robson 1983). Its properties of inheritance
and data encapsulation make it appropriate for our formalism (Ferber 1991). The agents
are organized into a hierarchical structure of classes according to their characteristics,
from the most general to the most specific. For instance, a town inherits all the properties
of settlements and these in turn inherit all the properties of places.
Multi-agent models use an anthropomorphic vocabulary that is not well suited to our
approach to settlement systems at a meso-scale, but which will be used for convenience.
What is called here the behaviour of a place or a settlement is in fact the aggregated
result of the individual actions of its inhabitants. Each settlement is said to be able to
accomplish a set of actions: for example, to consume, to produce or to exchange as if it
had autonomous behaviour.
79
The action space in which the agents operate is represented by grids of variable size
and shape. Each cell of the grid is called a place. Places can be of various shapes (even
irregular) but the most commonly used are squares and hexagons. Places are themselves
agents, as they need to communicate to be able to deal with events or information which
do not depend exclusively on settlements (e.g. the building of a road, or a natural
catastrophe). Each place has a type of natural environment (e.g. a plain, mountain, sea,
swamp, etc.) and may contain one or more parts of a network (e.g. a river, road, etc.). It
also receives natural (i.e. agricultural, mineral, maritime) resources, according to its
nature and position. These resources correspond to a potential to be exploited by the
population, the productivity varying with respect to such factors as technological ability
or the activity of neighbouring settlements.
Each inhabited place contains a settlement agent. Settlements are characterized by the
size of their population, their wealth and the functions they possess (agricultural,
economic, industrial, administrative, etc.). The nature of each agent determines its class.
For instance, a settlement which possesses only the agricultural function is a hamlet;
another with several economic functions is a town, its level depending on its size and the
nature of these functions. We assign a part of the population to each function, although
hamlets devote the whole of their population to agriculture.
The dynamics are simulated according to a set of rules which define the interactions
between agents as well as the constraints imposed by the environment. Some of the rules
are deterministic, depending only on constraints imposed by the environment, and others
involve random processes that produce a variety of trajectories. Thus two places,
although belonging to the same hierarchical level and having the same environment and a
similar productivity rate, would not necessarily have the same evolution.
Every simulation proceeds by a sequence of iterations. At the end of each iteration, a
balance of the exchanges and the actions performed during the period is made for each
settlement. The results modify the settlements for the next step through growth, the gain
or loss of functions, environmental changes and so on.
The population P of a settlement is distributed between the different activity sectors
according to the functions the settlement possesses. We have the relationship
where F constitutes the functions of a particular settlement. Each time a city acquires a
new function, a proportion of its population is devoted to it. This is done by transferring
fractions of the total population from old activities to new ones.
Settlements have an evolving demand for consuming various goods and services. The
total demand is a function of the total population P, and is distributed among the different
goods and services which are proposed. Let f be the propensity of a settlement to
consume the goods and services associated with the function f. The total demand D of a
place is:
Artificial societies
80
A settlement usually produces and supplies the maximum that its status allows. Each
subpopulation associated with an economic function f has a productivity f which
determines the value of its supply Sf:
Sf=f Pf
The places communicate the amounts of their supplies and demands by means of
messages. A place first satisfies its own demand with its own production. It can exchange
goods and services with other places according to its economic functions and the amount
of its remaining stocks. Productivity, propensity to consume and a range which represents
the maximal distance at which the goods are to be exchanged are associated with each
economic sector. Trade consists of the transfer of wealth from one entity to another. The
total supply of a commercial city is divided between the neighbouring units that are able
to buy some of it, according to a function of their wealth and of the distance between
them.
The adjustment between supply and demand is carried out iteratively by means of a
large number of complex interactions between places. For each cell, we calculate the
consumption of goods, i.e. the demand that is satisfied effectively. The price for each
good is proportional to the distance between the buyer and the seller. For each sector, we
compute an income, which represents the value of the stock of goods that has been sold.
The growth of a place depends on the previous trading balance, and particularly, on
the actual profits. According to whether the balance of trade is negative or positive, the
growth rate is set above or below the average growth rate of the system. More precisely,
it is the result of a normally distributed random variable.
The relative growth rate of the subpopulation of each economic sector depends on the
amount of unsold goods in that sector. If a place has some goods left unsold in one sector,
this means that its population for that sector has grown too fast or that the market has
shrunk. As a consequence, the share of this activity in the total population will be
reduced. On the other hand, if a sector makes profit and its market is growing, we
increase its population according to a normal random variable. If the share of the
population devoted to a given function drops to zero, the settlement loses this function.
Each settlement then is tested to see if it can acquire new functions. The criteria to
meet (different for each function) usually depend on size, wealth and which functions the
settlement already possesses, and this can involve such factors as the existence of the
function in a neighbouring settlement, or be partly related to a random process.
This short explanation of the model shows the simplicity of its inner working. All the
complexity goes into the handling of the multiple local interactions between the various
components of the system. The framework is very user friendly, because the stages of
modelling follow very closely the steps of geographical reasoning mentioned above.
A simple example
81
To illustrate this description of the model, we can observe the results of a specific
simulation. The parameters for this simulation are:
1. The environment is a hexagonal grid of about four hundred cells.
2. The length of a cycle is ten years.
3. The initial distribution of resources and population is random over a small range of
possible values.
4. There are only three functions available:
Artificial societies
82
Even with so few constraints, after the simulated passage of 1000 years, an embryonic
hierarchy is clearly visible on the screen (see Figure 5.1). Each settlement is represented
with a size proportional to its total population using a symbol corresponding to its
functions. Empty hexagons are uninhabited (swamps or mountains).
83
more so, the higher their level. However, they did not all grow at the same rate because of
the competitive market (larger and nearer cities products are more attractive).
A rank-size representation of the settlement-size distributions over time shows that the
process of hierarchicalization is very slow (see Figure 5.4).
Figure 5.5 plots the slopes of the rank-size curve during the simulation. It was noticed
from previous simulations that a permanent rate of innova-
Artificial societies
84
Concluding remarks
The application of tools and methods derived from AI to geography, although very
promising (Couclelis 1986), is still in its infancy. We have shown how multi-agent
systems could help to build simulation models for studying the evolution of complex
geographical systems. They are more flexible and easier to handle than the usual
mathematical formalisms and can be built step by step.
This method of modelling is a way to handle the problem of dealing with the double
character of the evolution of human settlements from the perspective of global change.
On the one hand, rather strong deterministic and general processes lead to a remarkable
convergence in the properties of settlement systems all over the planet. On the other
hand, diverse local conditions lead to a specific spatial organization in each autonomous
territory and confer a unique historical trajectory to every village, town or city.
Further developments of this model should allow us to test the major current questions
involved in describing settlement systems as self-organized: is it possible to identify
which local properties and interactions, defined at the meso-level of spatial aggregates,
can produce the global dynamics that lead from a scattered rural pattern of settlements to
85
a more concentrated form of human habitat? Is such an evolution, which totally changed
the relationship of societies with their environment, still reversible, and if so, under what
conditions?
In order to answer such questions, we would also like to use the model to test
hypotheses about the driving forces which may lead to different spatial configurations of
settlement systems all over the world. We think that two parameters are decisive in
controlling the spatial arrangement and degree of hierarchicalization of settlement
systems: the average speed of communication which prevails at the moment when the
settlement system starts its major bifurcation; and the relative importance of the bottomup and top-down processes generating an urban organization. That is why we want to test
the influence of these two factors on the development and characteristics of settlement
systems by means of our simulation model.
Chapter 6
The EOS project: integrating two models of
Palaeolithic social change
Jim Doran and Mike Palmer
A previous report on the EOS project (Doran et al, 1994) described the development of a
specialized software testbed, and the design, implementation and testing within it of a
simple computational version of the Mellars model for the growth of unprecedented
social complexity in the Upper Palaeolithic period of south-western Europe (Mellars
1985).
In this chapter we describe new EOS project work and new results, which additionally
address Gambles (1991) alternative model of social change in the Upper Palaeolithic
period. We also discuss some important methodological issues that have arisen.
87
Artificial societies
88
(c) a variety of adjustable parameters including those that specify the behaviour of the
resources.
89
reported we have instead linked co-operation to the agents need to avoid clobbering.
By clobbering we mean that one agent may accidentally damage the execution of
anothers plan (e.g. when two agents plan independently to acquire the same resource,
and when only one can be successful).
Artificial societies
90
intuitively there would be a much greater relative immobility and territoriality amongst
agents. We must therefore think a little differently about the relationship between a unit
of distance in the testbed and distance in antiquity. And, of course, we now need an
interpretation of alliance within the model.
EOS2
We now describe and explain our new extended computational model, EOS2. Most
extensions have been implemented by adding new rules to, or modifying existing rules in,
the standard production rule set associated with agents (with its supporting function
library) and by simple extensions to the generic social model and other representational
structures within agents.
Agent planning
Agents in EOS2 must be able to plan the co-ordination of many agents, not to acquire
complex resources (as in EOS1) but rather to avoid inter-agent clobbering. This requires
them to undertake a more complex form of planning than in EOS1. Agents decide the
resources to target and the number of agents needed to acquire them, and then seek to
make an agreed allocation of agents to resources. A single agent and its following may be
assigned several resources to harvest. Plans have explicit estimated numerical payoffs.
The way in which the anticipated payoff of a plan is calculated takes the risk of
clobbering into account (among other things) so that agents can see the advantage of
co-ordination.
It is important to keep in mind that agents are planning in terms of what they believe
to be true, rather than what in fact is true. Although agents awareness of resources and
other agents is accurate (if possibly incomplete) in all the experiments we describe in this
chapter, the judgements that agents make are heuristic and not tightly related to the
reality of the agents world. In particular, an agent may wrongly believe that one plan is
likely to be more rewarding than another, or may believe that a proposed plan is
sufficient to meet its needs when it is not, and vice versa.
Plans are built incrementally by an agent, thus:
WHILE
(a) the estimated payoff of a plan to the planning agent is less than the agents believed
energy level deficit;
(b) the size of the plan is less than the maximum plan size for the agent;
(c) adding a new resource will increase the estimated payoff of the plan to the planning
agent; and
(d) there is a new resource which can be added
DO Add an appropriate new resource to the plan.
There is a distinction between the total expected payoff of a plan and the expected
payoff of a plan to the planning agent. The latter takes into account the need to share out
acquired resources amongst all the agents involved (which may or may not include
followers and allies of the planning agentsee later in this chapter). Factor (a) captures a
91
notion of sufficiency: an agent stops planning if it has a plan which looks as if it will
suffice to meet its current energy deficit. Factor (b) captures a notion of bounded
rationality. Different agents have different limits on the size of plan they can build.
Within factor (c) a balance is struck between the assumption that the smaller the plan, the
more vulnerable it is to clobbering, and the larger the plan the more generally
unreliable it is. Factor (d) requires the agent to refer to its model of resource availability.
In general this is incomplete and is built up both by the agents own observation and by
information coming from any followers the agent may have.
In a resource rich environment, the agents will typically believe that small
individual plans suffice. However, as the number of resources and their richness
decline relative to the number of agents, small plans will be judged to be less rewarding
and no longer sufficient (both because individual resources are less rich and because the
likelihood of clobbering increases) so that agents will tend to build and execute more
complex plans. That in turn means that the bounded rationality limit on the size of plan
an agent can build will become much more significant. Notice that this planning
mechanism embodies a notion of economic rationality (e.g. Doyle, 1992).
Plan adoption and execution: the beginning of co-operation
In principle, agents all plan concurrently and asynchronously. Which agents plans are
adopted for execution? This is a complex matter. In EOS2, agents select and invite other
agents to join the plans they have created (selecting first their own followers). Agents
adopt the plans that they judge to be potentially the most beneficial to themselves in
terms of their own current beliefs: either they persist with their own plan, or they agree to
join that of another agent. The effect is that, with some delay, the more highly rated plans
are wholly or partially adopted for execution by groups of agents. One of the agents in
each group is the originator of the plan and is therefore viewed by the others in the group
as potentially a leader.
Plan execution is also complex. It involves the setting of a rendezvous point and time,
and the pooling and sharing at the rendezvous of all the resources harvested. Further
complexity is created by the possible failure of some or all of the agents to perform the
tasks they have agreed to undertake, and the fact that the whole process is partly
recursive: a subleader will execute its part of a plan by creating and recruiting to a
subplan and so on.
The properties of this co-ordination process (which involves a form of team
formation, cf. Kinny et al. 1992) are complex and difficult to establish theoretically. The
process is not intended to be in any sense optimalrather, it is intended to be effective
and plausible in its broad characteristics.
Alliances in EOS2
Recall that any relationship instance is expressed by beliefs within the social models of
the agents concerned. For example, to say that agents X and Y are in a leader-follower
relationship means that Agent X believes that Agent Y is its leader, and vice versa.
In EOS1 such a leader-follower relationship came into existence whenever one agent
(which might itself already be a leader) agreed to join anothers plan. This is
Artificial societies
92
93
(i) tendency to be greedy or altruistic i.e. the agent proposing the plan may seek to
retain more or less than its proportionate share of the proceeds.
Factors determining the claimed quality of the plans an agent proposes:
Results
We indicated earlier the collective behaviour that the Mellars (1985) and Gamble (1991)
formulations suggest: as the number and richness of the resources available to the
population decline without resource concentration and with unpredictability of individual
resources, so (following Gamble) the number of inter-agent alliances should increase. But
if there is resource decline coupled with survival of individually, stable resources in
limited areas then (following Mellars) population concentration and development of
complex hierarchies is predicted. In fact, our experimental findings are only partially in
line with these predictions.
We find that many factors bear upon the dynamics of the agent community and the
alliances and hierarchies within it. Three of the most important are:
1. Resource availability. Can the agents acquire sufficient resources to survive? This
involves the maximum rate at which agents can harvest, as well as the actual rate of
resource renewal. Inadequate uptake rate means that agents must die.
2. Resource dispersion. The more spatially dispersed the resources are, the more difficult
it is for agents to co-ordinate.
3. Agents hunger levels. Is the present energy level of the agents up to their target
level? If so, they will construct minimal plans and no co-ordination will occur. If not,
they tend to plan more because agents believe that more complex plans avoid
clobbering and hence yield a better harvest.
Recall that in all the experimental trials reported here, agents create and work with
accurate, but possibly incomplete, social models. Additionally, all resources in these
experiments are simple in the sense that each can be harvested by a single agent.
Resources may differ, of course, in their energy yield when harvested.
We have adopted simple measures of alliance and hierarchy. The alliance measure is
simply the number of alliances (between pairs of agents) divided by the total number of
Artificial societies
94
agents. The hierarchy measure is the sum of the hierarchy depths of the agents in the
hierarchy, again divided by the total number of agents. A leader has a hierarchy depth of
zero, one of its followers has depth one, a followers follower has depth two, and so on.
The depth of a hierarchy is defined as the maximum agent depth it contains. In graph
theoretic terms, a hierarchy is a tree rooted at its leader.
A low resource period trial
The impact of two of the factors just listed, resource availability and agent hunger
levels, on hierarchy and alliance formation is illustrated by the following experimental
trial. A group of 12 agents is subjected to a resource regime consisting of an initial period
of ample available resource energy, a comparatively long period of very low available
resource energy and then sustained ample resource energy again (see Table 6.1).
The resources are spatially semi-distributed: randomly scattered over an area of
500500 distance units within a 10001000 environment (see Figure 6.1).
Resources renew every 30 cycles. Agents consume one energy unit each per cycle.
They begin with energy levels of 1000 and with a target energy level of 1000. Agents
awareness range is set at 200 units of distance
Resources
1200
2011200
12012500
95
Artificial societies
96
and co-ordination structures appear (i.e. alliances or hierarchies), but then (at around
12001500 cycles) deaths
98
Artificial societies
99
Artificial societies
100
Table 6.2 shows the results of 322 trials, each involving two replications with two
different pseudo-random number streams. For each replication of each trial, results are
presented after 1500 cycles: the number of agents surviving, the values of the alliance
and hierarchy measures, and the average energy level of the surviving agents. The trials
vary three factors:
Varying resources
(VR)
VA+VR
Intermediate
Al
Hi
AE
10
1.4
1.1
463
0.6
1.6
0.4
Ag
Dispersed
Al
Hi
AE
0.2
0.8
294
354
0.2
0.8
1.5
524
0.3
0.1
2.6
242
11
0.5
1.3
384
0.0
1.7
0.2
12
1.0
Ag
Al
Hi
AE
0.0
0.6
271
243
0.0
0.4
524
1.2
249
0.0
0.7
82
0.0
0.7
262
0.2
0.7
398
0.0
1.2
237
0.2
0.7
463
334
0.0
1.2
257
0.0
0.5
302
2.2
398
0.0
1.0
233
0.0
0.5
435
0.4
397
0.0
0.6
380
0.0
0.5
322
Key: Ag, agents surviving; Al, measure of alliances; Hi, measure of hierarchies; AE, average
energy of surviving agents. All results after 1500 cycles.
101
Artificial societies
102
Finally, we note that there is good reason to believe that consideration of economic
rationality, however comprehensively captured within the model, is not enough. It is
intuitively apparent that being a leader has to do with the dynamics of emotional
energy (Randall Collins phrase) as well as good decisions. A number of authors (e.g.
Alexander 1989, Collins 1990, Hammond 1990) have considered recently the role played
by emotions in society, including the emergence of stratification. To simulate aspects of
the flow of emotional energy in a model of the EOS type would certainly be possible and
might suggest rather different conditions for the emergence of leadership and hierarchies.
Methodological issues
The EOS project raises seemingly important methodological questions. Science is about
finding useful and tractable models, whether mental, informal or formal. The attraction of
DAI modelling in this and similar contexts is easy to explain. It enables cognition to be
explicit in the model (cf. treatment of endomorphy in agents in Zeiglers (1990)
terminology) and encourages the use of symbolic rather than solely numerical
representation. But, as always, models have to be developed in a valid way if their results
are to be trusted.
A prominent feature of both the EOS1 and EOS2 models (and indeed of all DAI-based
models) is the large number of adjustable parameters, numerical and symbolic, embedded
within them. As these parameters are varied the behaviour of the model system varies,
sometimes dramatically. Getting interesting behaviours, or a clear understanding of
particular types of behaviour, is rarely easy. We feel that a number of important questions
arise from this observation.
1. Validity How can any model be taken seriously when there are so many arbitrary
parameters within it? Cannot any desired behaviour be obtained from a model with a
sufficient number of adjustable parameters?
2. Classification of behaviour How can the set of model behaviours be classified?
3. Abstractness Why should any one particular model be used when there are clearly a
range of alternatives to hand at differing levels of abstraction?
4. Formality Is there a need for formal support for the work?
We now attempt to give partial answers to these questions.
The validity of the EOS models
The EOS models have not been, and clearly cannot be, validated in anything like the
traditional sense of the word. The many assumptions within the models have not been
supported empirically, some of the claims made explicitly in connection with the agents
are controversial even within the DAI community (e.g. our reference to agents
beliefs), and the actual behaviour of the models has not been and cannot be checked
against reality in any detailed way.
But this lack of validity in the traditional sense is not as damning as might at first
appear. What the EOS models do is to give relatively precise interpretations to the
informal models of Mellars (1985) and Gamble (1991) and enable relatively precise study
103
of their consequences. That is an important step forward. Both Mellars and Gamble (and
others) are proposing possible processes which they feel may fit the limited available
evidence. Their emphasis is on insight and plausibility, rather than a precise fit to
particular events in antiquity. So is ours. There is an important sense in which our work,
like that of Mellars and Gamble, is more theory building than it is modelling. The aim is
to discover more about certain processes and how they interact. By so doing, the way is
prepared for more traditional modelling studies in the future. Yet at the same time we do
not exclude the possibility that testable hypotheses may be derived directly, as
experimental exploration focuses our attention on particular relationships between
structures and parameters.
It is not the case that any behaviour may be obtained where there are a multitude of
structural parameters. But in such circumstances experimentation must go beyond a small
number of trials, however interesting, to the formulation of a coherent and sound
overview of the possible behaviours that may be obtained from the model, and in what
circumstances they occur.
Understanding a models behaviour
Even when the usefulness of models of the EOS type is accepted in principle, there
remains the problem of how best to comprehend and to express so much possible
variation. It is not clear whether the behaviours of these models may be relied upon to fall
into simple patterns (as seems implicit in the informal models from which they are
derived) or whether we are faced with something more like chaosa multiplicity of
structural parameters, with unpredictable sensitivities and no simple patterns of behaviour
to be discovered. Should the latter be the case, progress may be very difficult indeed.
Choosing between differing levels of abstraction
A further issue is how best to make the choice between the many alternative models, at
varying levels of abstraction, whose behaviour might be studied to address a particular
problem. For example, we may imagine a model, called ABSTRACT-EOS, whose
operation is as follows:
FIRST
scatter resources over a two-dimensional region according to a multi-normal
distribution and agents over the landscape according to a uniform distribution
THEN REPEATEDLY
compute a (locally) optimal multi-agent/multi-resource plan for each agent based upon
the resources and other agents within the agents sensory range (factoring in any
existing alliances and leader-follower relationships);
compute the set of plans to be executed as that subset of the agent plans which satisfies
the constraint that no agent should appear in more than one plan in the set, and which
provides the maximum total payoff;
Artificial societies
104
compute the payoff from the plan set to be executed, and discard/ replace any agents
that die (note that clobbering may occur); and
insert and discard examples of alliances and dominance relationships by simple rules of
the type used in EOS2,
WHILE monitoring the dynamic pattern of alliances and leader-follower relationships.
Using this model, we might expect to demonstrate that agents located in the midst of
many resources become high leaders, but with the results sensitive to the size of the
agents sensory range and the parameters of the resource distributions.
The important feature of ABSTRACT-EOS is that it bypasses both the internal
cognition of agents and inter-agent communication, while retaining other essentials of the
EOS model. But it loses the important potential match or mismatch between agents
social models and social reality. The questions are, which is the more useful model, EOS
or ABSTRACT-EOS, and why?
A standard modelling principle is that the complexity and level of a model should be
chosen so that it answers the questions and embodies the theoretical elements we are
interested in, but is otherwise as simple as possible. Complexity for its own sake,
including cognitive complexity, is to be avoided. Agents in a model should not, for
example, be programmed to do predictive planning if that complexity is not clearly
needed for the modelling objectives to be achieved. Does that principle apply here, and if
so, how is the judgement of need to be made? It might be argued, alternatively, that when
modelling human society we should focus on the human level, and regard higher
phenomena as emergent from that level, if only because that makes everything easier for
us to understand. We feel that this issue is unresolved.
A need for formal support?
It may be suggested that the difficulties with the EOS models just discussed demonstrate
the need for some more formal framework within which these models should be set (or
by which they should be replaced). This argument will appeal particularly to those
researchers within distributed AI studies who are taking a more logicist road to the study
of teams and societies (e.g. Kinny et al. 1992, Werner 1990).
The argument is attractive but is, we believe, less persuasive than it at first appears.
Such a formal framework typically will be a system of axioms and inference rules, or
some partially formulated version thereof. But how can this help? Certainly, certain types
of general inference could be made and abstract theorems proved. But in the context of
human society, a formal framework will always be either a gross oversimplification (at
best it will illuminate certain essentials) or replete with domain specific axioms and
tractable in particulars only as a computer simulation of essentially the EOS type.
Unquestionably, any model should be designed in a principled way. But we must
avoid imposing additional mathematical requirements which in practice compel a model
to be too simple to be of value and which may even focus attention away from the
essential requirements of the modelling task.
105
Conclusions
We feel that we have already achieved some useful and original insight into the close
relationship between the Mellars and Gamble models. We find that our experimental
results, as far as they go, are compatible with the Mellars (1985) view but not yet with
(our interpretation of) Gambles (1991). But much more experimental work is needed.
And there are important methodological issues that have yet to be fully clarified and
resolved.
Acknowledgements
We are grateful to our colleagues Nigel Gilbert and Paul Mellars for many valuable
discussions related to this chapter, and to Nigel Gilbert for making available additional
computing facilities at the University of Surrey. The EOS project has been supported by
the UK Joint Council Initiative in Cognitive Science/HCI by grant no. SPG 8930879.
Chapter 7
Genetic algorithms, teleological
conservatism, and the emergence of
optimal demand relations: the case of
learning-by-consuming
Roger McCain
The demand relationship (the relationship between the price of a good and the quantity of
that good demanded) has long been a central construct of economic theory. Economic
theory conventionally assumes that consumer demand curves are determined by the
maximization of utility or preference. That is, the consumer demand relationship is the
locus of price and quantity pairs such that the quantity demanded yields maximum utility
or preference ranking at the given price.1 However, a large body of evidence from
cognitive science, experimental economics and management studies indicates that people
are not capable of judging maxima in many conditions.2 Instead, rationality consists of
behaviour according to heuristic rules which, while not optimal, give satisfactory results
in a wide range of cases.3
The consumer demand relationship might instead be considered as an heuristic rule, or
as the predictable outcome of a set of heuristic rules. This would be consistent with
modern cognitive science and with a number of proposals for a more cognitively realistic
conception of economic rationality. Among the many questions that then arise are: (a)
How do these heuristic rules arise and what determines them? If they are not hardwired
into the human genetic constitution, then their origin and characteristics should
themselves be the subject of investigation; (b) Can we characterize with any generality
the heuristic decision processes that consumers use? and (c) Is it possible that these
heuristic processes lead, in some cases, to approximately optimal results? If so, how good
is the approximation, and what determines its quality from case to case? Of course, if it
should prove that the answer to questions (b) and (c) is yes, and that the approximation
is almost always quite good, we could then return to the neo-classical optimality theory
without further concern about the specific cognitive processes involved.
One obvious possibility is that the heuristic rules evolve in some sense (Nelson and
Winter 1982). However, evolution is not a uniquely specified process. A recent
development in computer science, genetic algorithms (Holland 1975, Goldberg 1989,
Koza 1992) provides a model of evolution that includes an analogy with sexual
reproduction.4
This chapter reports on computer simulations of the determination of consumer
demand by such heuristic processes. The chapter begins with a relatively conventional
demand relationship and then proceeds to apply these processes to some more
107
complicated models of demand, including cultivated tastes. In each case the question is
whether, and to what extent, the simulated agents learn approximately optimal behaviour.
Genetic algorithms
The simulations reported in the previous section were based on models that share a
common shortcoming which is the converse of the criticism of neo-classical economics:
they assume cognitive processes that are more rudimentary than real human cognitive
processes. The truth presumably must be somewhere between these models and neoclassical maximization.
Suppose that demand rules evolve (Nelson & Winter 1982, McCain 1992b,
forthcoming). That does not specify a single sort of model for theoretical work or
simulation. Evolution is a process of random change and selection. Then what varies, and
in what sense does it vary randomly? And how does selection come about? These are by
no means trivial questions and it may be that seemingly arbitrary answers to them
determine the outcomes of a simulation.7 A recent literature that shows some promise in
this connection is organized around the idea of genetic algorithms. A genetic algorithm is
an algorithm based on random variation and selection in which the random variation in
the main takes the form of random recombination8 of given elements, rather than
mutation. A large literature now exists on the use of such algorithms to find numerical
maxima, fit curves, and evolve highly effective rules of procedure (Holland 1975,
Goldberg 1989, Rawlins 1991).
The routines in the simulations reported here are demand rules; that is, rules that make
the quantity purchased a function of price.9 Both the mathematical specification and the
Artificial societies
108
parameters of the demand rules are coded as inputs to a genetic algorithm. The details of
this coding are reserved for the Appendix on page 138. The population is seeded with
demand rules with randomly determined parameters. Selection to reproduce is based on
current-period utility (rather than, for example, average utility over a longer period),
when the rule is applied in the context of a price that is a stationary random variable.
Thus, the demand routines of a simulated individual consumer are modelled by a
string of about 160170 zeros and ones, the consumers chromosome in the genetic
algorithm. On each step of the simulation, the population (of demand routines) is
reconstructed as follows. First, two simulated consumers demand routines are selected to
reproduce. The mates are selected by a pseudo-random routine that assigns a higher
probability to the selection of routines that yield a higher utility. The chromosomes are
then split at a randomly selected point and the first portion of one linked to the last
portion of the other, and vice versa.10 (This simulates the biological phenomenon of
crossover in sexual reproduction, according to some accounts of the rationale of
genetic algorithms.) This is repeated until there are 100 members of the new population.
Before this recombination takes place, mutations occur at a 2 per cent rate11; that is, two
of the 100 strings are chosen at random and one of the zeros is changed to a one, or a one
to a zero, at random.
Since these programs are meant to simulate the behaviour of consumers, we should be
explicit as to the behaviour they are meant to simulate. It is emulation.12 The idea is that
consumers reconsider their demand routines periodically and adoptin partthe
routines of other consumers who have had greater success in matching their purchases to
their needs. Thus the more successful routines would spread through the human
population by a process of imperfect emulation.
The question, then, is this: how effective can such a process be in approximating
optimal demand? The genetic algorithm seeks the optimum quantity demanded fairly
effectively. Figure 7.1 gives an indication: it plots the simulated quantity demanded for
100 steps of a representative simulation run in which price is constant and the optimal
quantity demanded is always 980.13
The performance is not especially impressive by comparison with simpler heuristics
(McCain, forthcoming); but one would think that the heuristic demand process would
have the advantage when price, and thus optimal quantity, varies, since in a case of that
kind the heuristic rule would be capable of responsive behaviour.
109
Artificial societies
110
111
Teleological conservatism
In earlier work, McCain (1991a, forthcoming) explored similar models in which choices
are determined by simple groping, that is, a process of random trial and error in which
trials that lead to deterioration of the results are abandoned. The genetic algorithm,
however, is not a generalization of groping. The simple groping model has a property the
genetic algorithm lacks that makes a difference to the process of selection, a property we
might call teleological conservatism. In a genetic algorithm (and in evolution)
reproduction is, as it were, a pure lottery. The parent organisms, or code, are lost.15 The
filial organisms, or codes, are selected from more fit parents, but because the code is
recombined randomly, there is always some probability that the filial organism, or code,
will be less fit than the parents. In a groping process, by contrast, the performance of the
recombined heuristic is compared with that of the parent heuristics, and if the filial
heuristic is less effective, it is discarded and the parent heuristics kept. Thus groping
selection is not a lottery, but an option. To put it in more biological terms, it is as if the
parent, having engendered a child, could reject the child if it is less fit and replace the
child with a clone of the parent itself. Of course, evolution by mutation and sexual
reproduction and selection of the fittest also lacks the property, as we might suspect,
since the genetic algorithm is meant to model the evolutionary process. Thus the property
of teleological conservatism serves to distinguish groping from evolution.
An advantage of genetic algorithms is that they can find the neighbour-hood of the
global optimum in many optimization problems in which there are multiple local optima
(Rawlins 1991). It is not clear that a groping process with teleological conservatism
would have this capacity. Biological evolution has given rise to manifold creativity in
natural history, but a groping process may not be able to do the same.
However, our purpose here is not to model evolution or to maximize a function but to
model the determination of demand, when people are imperfectly rational but are capable
of learning by emulation of one another. The process is as follows. Each person
determines his or her own demand by means of some heuristic routines that would cut
back consumption when prices rise. A person then observes the success of some other
consumer, and selects a relatively successful one to emulate. Success is measured in
terms of utility achieved, but since utility is imperfectly perceived, the person chosen for
emulation may not be the single most successful one. The emulation of the chosen
models routines is also imperfect, and takes the form of random recombinations of the
models and the imitators own demand routines. These recombinations form a new
population of demand routines, which replace the old ones only to the extent that the
replacement increases the utility of the imitator and which are rejected otherwise. If they
could not be rejectedif the new population of routines were to replace the old,
regardlessthen we would have a genetic algorithm.
In what follows, genetic algorithms (without the property of teleological conservatism)
will be called standard GAs, or just GAs, and modified genetic-like algorithms with the
property of teleological conservatism will be called heuristic groping or teleologically
conservative algorithms (TCAs).
We apply this procedure first to a conventional demand example. The results will be
summarized very briefly. It appears that in these models, either GAs or TCAs can give
rise to approximately optimal demand relations, predictably and reliably. One might
Artificial societies
112
suppose that teleological conservatism could be helpful when individuals have differing
utility functions. The danger of a standard GA in such a case is that the individual would
emulate (borrow code for an heuristic routine) from another individual whose utility
function is quite different from her/his own. The success of the individual emulated
would not signal that the same routines would be successful for the consumer who
emulates. To investigate this possibility, simulations were run based on linear demand,
with one modification: utility functions are allowed to differ among the simulated agents.
Again, the results will be summarized very briefly. Here, as before, the aggregate market
behaviour of the simulated consumers yielded an average demand rela-tion of the
conventional type, and in the rough neighbourhood of the optimal average demand, after
some delay as optimal demand patterns were learned. This was true both of the
standard genetic algorithms and of the TCAs. However, the TCAs gave rise to a better
average performance.
Learning-by-consuming
In a number of papers written in the later 1970s and early to middle 1980s (McCain 1979,
1981, 1982, 1986), I explored the implications of the cultivation of taste in otherwise
fairly conventional models of the demand for works of art. All of those models assumed
that consumers act as if they maximize a utility function, although in most of them 16
consumers are assumed to be short-sighted. A key prediction of my earlier, as-ifmaximization models is that the demand for works of art is multi-modal.17 We shall see
that this prediction extends also to some of the cognitively realistic models.
We shall model the demand for art. In the terms of my 1979 and subsequent papers,
the hypothesis of cultivation of taste is expressed in the following way: the consumers
utility depends on the flow of a subjective quantity we might call enjoyment-of-art. The
subjective quantity will be denoted by Q and the utility function used in the simulations
will be
U=+5Q0.01Q2
where represents spending on all other goods and services.
In turn, the subjective flow of enjoyment-of art depends on the objective quantity of
art consumed (Stigler and Becker 1977). Greater experience leads to greater sensitivity to
the benefits of art and thus to greater consumption and enjoyment. In economic
terminology, this experience is a kind of immaterial capital, which we might call taste
capital, and the learning process is the accumulation of taste capital. In this study the
objective quantity of art consumed will be denoted by X and the sensitivity by , and the
dependency of the subjective flow on the objective flow is
Q=X0.2
Thus is a measure of taste capital. If the consumer were to maximize utility, the optimal
demand relation (written as a demand price function) would be
p=5.0040.022X
In these simulations the taste parameter is determined recursively by
113
t=0.9t1+0.1Xt1+ 0.002
This might be called a learning-by-consuming model by analogy with learning-bydoing in the economics of technology and production (e.g. Arrow 1962).
In simulations of learning-by-consuming based on simple groping, the average
quantity demanded and utility attained differ systematically from one simulation to the
next (McCain, forthcoming). This is because random factors initially and during the
groping process lead a smaller or larger proportion of the population to develop
cultivated tastes. We are particularly concerned with the prediction that cultivation of
taste would lead to multi-modality of demand. These simulations show the bimodal
distribution of cultivation of taste previously predicted on the basis of catastrophe theory
(McCain 1979).
When learning-by-consuming is modelled in a simulation based on a standard GA, the
results are far less positive (McCain 1992b). Multi-modal outcomes are not observed.
Depending on the details of the simulation, these simulations could eliminate rapidly any
taste capital randomly assigned in the initialization process and converge rapidly to very
low rates of consumption and taste capital, or converge equally rapidly to universal
cultivation. Nor was optimal demand well approximated. Figure 7.4 shows a simulation
that led to universal cultivation.
In this simulation, the highest utilities were emulated, so that those who were
randomly assigned a high level of cultivation at the outset were more likely to be
emulated. The grey line shows the optimal demand for a culti-
Artificial societies
114
115
Artificial societies
116
117
Artificial societies
118
Thus there are two ways to code for a Cobb-Douglas function, i.e. the product of power
functions of the independent variables, a function often used in mathematical economics:
110011100011000 and 101001010010100.
Acknowledgements
I am indebted to participants in the Seventh International Conference on Cultural
Economics, Fort Worth, Texas, October 1992; the Seminar on the Economics of the Arts,
Venice, Italy, December 1992; and the Symposium on Simulating Societies, Siena, Italy,
July 1993; for useful comments on some of the research reported in this chapter. Errors
and omissions are, of course, attributable to the author.
Notes
1. It should be observed that this is a statement about the individual demand curve, not the
aggregate demand curve. Economic theory has sometimes posited the hypothesis that, if the
individual demand relationships are derived in this way, aggregate demand relationships will
have certain characteristics that are sufficient for a market equilibrium to exist. This may be
called the representative man hypothesis in demand theory. However, it is now clear that
utility-maximizing demand relationships are neither necessary nor sufficient for the
representative man hypothesis, and thus for the existence of market equilibrium.
Nevertheless, it is basic to many applications of the demand concept, for example, in costbenefit analysis, and conversely, evidence of responsiveness to price-like variables is often
offered in the economic literature as evidence of the rationality of economic action. These
facts justify a continued interest in the rationality of individual demand relationships.
Moreover, issues with respect to the aggregation of demand relationships are raised, not
resolved, by the finding that utility-maximization-derived individual demand functions are
neither necessary nor sufficient for the representative man hypothesis. Aggregate demand
relationships as described in the representative man hypothesis might emerge in any case,
not as necessary but as contingent outcomes of individual decisions, at least in some
circumstances. This would constitute emergence, and is precisely the issue to which this
chapter is directed. In order to address it, it is necessary to avoid both of the contradictory a
priori hypotheses about demand relationships common in the social science literature: the
neo-classical hypothesis of utility maximization and the hypothesis that, since people are
obviously not rational, demand decisions have nothing to do with price and utility
maximization considerations. The truth may be in between the two, or partake of both
hypotheses, and adopting either would then blind us to the important issue of when one or
the other is reliable. It should also be noted that this chapter is limited to the preliminary
issue of the rationality and aggregateability of demand, and does not address the existence or
characteristics of market equilibrium, issues which are confronted by the work of Chattoe
(1993).
2. For example, Camerer (1987), Kahneman and Tversky (1984), and Kahneman, Slovic,
Tversky (1982).
3. As Herbert Simon (1981) has long argued. Note also Reddaway (1936) for a very early
example. For a survey see McCain (1992a).
4. By contrast, such earlier studies as Nelson & Winter (1982) allowed only mutation.
5. That is, a good in terms of which prices and incomes are measured. For purposes of this
discussion, money is a veil and plays no part in the model.
119
6. These are statistical processes commonly assumed, and often observed, in econometric work.
Constant prices are of limited interest to us, since a major concern of this chapter is with the
responsiveness of the simulated agents to changing prices. Thus, constant price assumptions
play only a comparative role.
7. Nelson & Winter (1982), for example, find that they have to make use of the production
function as part of the specification of their simulation of evolution in firms. This seems to
be something of a failurewhat they give us is not a model of evolution but of adaptation.
8. I cannot resist the temptation to observe that in the Theory of economic develop-ment,
Schumpeter habitually refers to innovations as new combinations.
9. This, of course, imposes some unrealism by identifying the heuristic with the demand
relationship itself. There would be merits in a more realistic heuristic, and future work
should address that. However, there are pitfalls in such a research strategy. Suppose, for
example, that a model were to identify a budget of total spending on a commodity as the
demand heuristic that is to evolve. This might be more realistic, but it entails a specific
demand pricequantity relationship, and one that is highly responsive to price changesthe
implied relationship would have an elasticity of exactly -1. The assumptions made here are
in fact less restrictive and allow for unresponsive demand heuristics, the very possibility we
wish to confront.
10. One implication of this is that the functional form is unlikely to be changed by
recombination, since the probability that the split takes place in the first 15 of 160 or so
places is small. Thus it may be more accurate to say that the functional form is selected in
this study than that it evolves. One way to force the functional form to evolve would be
to let the first 15 positions play the role of a second chromosome, and have it recombined
on every round of the simulation. Multiple-chromosome models have been proposed but
have not been explored extensively in the genetic algorithms literature.
11. This rather high rate was chosen without any particular rationale, but some experiments with
alternative rates suggested that the rate of mutation made little difference. For present
purposes, mutation is a cognitive phenomenon, representing incorrect memory, perhaps.
Thus it is not clear what rates might be realistic, but we may say that, in a certain sense,
higher rates correspond to lower cognitive capacity.
12. For a very deep and useful discussion of the rationality of imitation or emulation, see Pingle
(1992). Pingles paper, excellent as it is, is somewhat compromised by his adherence to the
neo-classical conception of rationality, a conception which is also used in this chapter. As I
pointed out in my unpublished comments on Pingles paper, his ideas would fit more
naturally in a relative concept of rationality, such as my linguistic conception (McCain,
1991b, 1991c, 1992a). A referee observes that people imitate behaviour, not heuristic
routines. Perhaps; and there might be much interest in a study in which individuals model the
behaviour of others and consider emulating the behaviour they model (Johnson-Laird 1983).
That is beyond the scope of this chapter. In any case, it does not seem clearly wrong to
suppose that people have a repertoire of heuristic routines which are a common social
heritage that they recombine in various ways as their heuristics evolve. There being some
connection between heuristics and behaviourthat, of course, is the whole point of
heuristics!this would lead to imperfect emulation of heuristics, which is what the chapter
assumes.
13. The simulation does not converge to as near this optimum as do the models of simple
groping (McCain 1994) and remains below it. It has been noted in other studies of genetic
algorithms that their convergence tends to be incomplete. Due to initialization conditions, the
approach to the optimal quantity is from below. It may be that these considerations together
explain the tendency of the quantity demanded to remain below the optimal.
14. That is, the linear form requires that the constant and the coefficient be of very different
orders of magnitude, while the inverse specification did not. Thus, when the two constants
were initially random draws from the same pool, the inverse specifications tended to perform
Artificial societies
120
best in early rounds of the simulation, even when the true specification was linear. This
experience indicates that some functional forms are much more sensitive to initialization
conditions than others and thus may gain an early advantage in an evolutionary process that
prevents the correct specification from emerging. A different selective process, designed to
retain diversity in early rounds, might perform better.
15. I should remark that in practice, applied genetic algorithms have often included some
element of teleological conservatism. To that extent, however, they depart from the
evolutionary metaphor. This issue seems more important for simulation than for applications
to function maximizationin the latter case, the correct assumptions are the ones that work;
in simulation, the case is much less clear and needs explicit discussion.
16. The exception is McCain (1981).
17. This is testable econometrically. See McCain (1992a). I note in passing the econometric
support for a catastrophic long-run demand model for wine in my 1979 paper, which I
believe to be still the only implementation of catastrophe theory in econometrics.
18. In other words, this outcome presents an aggregation problem in which average demand is
not the demand of a representative consumer. Indeed, in a multimodally distributed
population, there can be no representative consumer.
Part II
The evolution and
emergence of societies
Chapter 8
Emergence in social simulation
Nigel Gilbert
Every discipline is based on a unique foundation of epistemological assumptions and
concepts. This means that even when one discipline develops so that it begins to share its
concerns with another, there may be little or no contact because the practitioners are,
literally, speaking different languages. Even if contact is established, the neighbouring
disciplines may still have nothing to say to each other because, while a topic may be
common, the questions being asked and what count as interesting answers differ so
greatly. To bridge the gap, considerable work in translation has to be done.
This situation seems to be developing in the case of artificial intelligence (AI) and the
much older discipline of sociology. Both are interested ostensibly in the behaviour,
structure and properties of collections of actors or agents. Yet the literature on distributed
artificial intelligence (DAI) makes little or no reference to sociology, or even to its major
concepts. Similarly, there are very few sociologists who have considered the relevance of
DAI to their concerns.
In this chapter, I shall show first that sociologists have also struggled with one of the
basic conceptual problems that those interested in simulating societies have encountered:
the problem of understanding emergence and, especially, the relationship between the
local and global properties of complex systems. Secondly, I shall indicate some ways in
which DAI simulations may have oversimplified important characteristics of specifically
human societies, because the actors (agents) in these societies are capable of reasoning,
and do so routinely, about the emergent properties of their own societies. This adds a
degree of reflexivity to action which is not present (for the most part) in societies made
up of simpler agents, and in particular is not a feature of current DAI simulations.
The macro and the micro level
The issues I shall address stem from the fact that both societies and the computational
systems with which distributed AI are concerned are composed of many interacting
agents (also known as people, actors or members). These systems can therefore be
described either in terms of the properties and behaviour of the agents, or in terms of the
system as a whole. The former mode of description focuses on the micro level, that is,
the features of individual agents and their local environment (which they can perceive
directly), while the latter focuses on the macro level, that is, the global patterns or
regularities formed by the behaviour of the agents taken as a whole. The problem is how
to characterize the relationship between these two types of description, in particular when
the macro properties can be said to emerge from the micro-level behaviour.
123
Methodological holism
The relationship between local and global properties is one which has exercised
sociologists since the foundation of the discipline. For example, Emile Durkheim, one of
its founding fathers, emphasized the external nature of social institutions (a macroproperty of a society) and argued that they imposed themselves on individuals at the
micro-level. It is worth quoting at some length from his methodological writings, because
he was addressing issues very similar to those which continue to worry people who
construct social simulations. After having defined a category of facts that he called
social facts, he wrote
Another proposition has beenvigorously disputedit is the one that
states that social phenomena are external to individuals The states of
the collective consciousness are of a different nature from the states of the
individual consciousness: they are representations of another kind. The
mentality of groups is not the mentality of individuals; it has its own
laws To understand the way in which a society conceives of itself and
the world that surrounds it, we must consider the nature of society, not the
nature of the individuals. (Durkheim 1895:656)
Thus, for Durkheim, there are social representations which can be examined
independently of the individuals that make up the society. These social facts are
a category of facts with very special characteristics: they consist of ways
of acting, thinking and feeling that are external to the individual and are
endowed with a coercive power by virtue of which they exercise control
over him. (Durkheim 1895:69)
This view of the relationship between an individual and society was later developed by
Parsons (1952) into a normative conception of order which emphasized the role of
internalized norms in ensuring the integration of society through shared values and
obligations. Subsequent theorists have extended this tradition, which has become known
as methodological holism (ONeill 1973). It asserts that the social behaviour of
individuals should be explained in terms of the positions or functions of these individuals
within a social system and the laws which govern it.
Methodological individualism
In contrast, methodological individualists see macro-phenomena accounted for by the
micro-level properties and behaviour of individuals. The individualists position demands
that all the concepts in social theory are analyzable in terms of the interests, activities,
etc. of individual members of society.
Artificial societies
124
125
of rules, although because they may be in different local environments, the actions
they take will differ.
A typical example of such a complex system is an array of cellular automata. Each cell
can take on one of a number of internal states. State changes occur according to the
results of the firing of simple condition-action rules dependent only on the states of
neighbouring cells.
Some complex systems are also adaptive. In such systems, agents either learn or are
subjected to a process of mutation and competitive selection, or both. In learning systems,
agents are able to modify their rules according to their previous success in reacting to
the environment. In systems involving competitive selection, agents rules are altered,
either at random or by using some specific learning algorithm, and then those which are
more successful are copied while the less successful are deleted from the system. An
important characteristic is that agents adapt within an environment in which other similar
agents are also adapting, so that changes in one agent may have consequences for the
environment and thus the success of other agents. The process of adaptation in which
many agents try to adapt simultaneously to one another has been called co-evolution
(Kauffman 1988).
Emergence in complex systems
Because complex systems, whether adaptive or not, consist of many agents, their
behaviour can be described either in terms of the actions of the individual agents or at the
level of the system as a whole. In some system states the global description may be very
simple (e.g. if the agents are either not interacting or interacting in repetitive cycles, the
global description might be that nothing is happening), or exceedingly complex (e.g. if
the agents are in complete disequilibrium). In some circumstances, however, it may be
possible to discover a concise description of the global state of the system. It is in these
latter circumstances that it becomes possible to talk about the emergence of regularities
at the global level.
An example relevant to social simulation can be found in the work of Nowak and
Latan (e.g. 1993) which aimed to model the effect of social influence. This draws on
Latans theory of social impact which states that the impact of a group of people on an
individuals opinion is a multiplicative function of the persuasiveness of the members of
the group, their social distance from the individual and the absolute number of the
members. For example, in a simulated world in which there are only two mutually
exclusive opinions, the opinion adopted by an individual is determined by the relative
impacts on the individual of those propounding the two opinions. At any moment in time,
an agents opinion is determined by the multiplicative rule that implements the theory of
social impact, and is depend-ent only on the opinions of other agents. The simulation fits
the definition of a complex system.
The outcome of Nowak and Latans simulations is that coherent clusters of opinion
emerge and remain in dynamic equilibrium over a wide range of assumptions for the
parameters of the model. The population does not move wholly to adopt one opinion, and
minority views are not completely expunged. Nowak and Latan show that the clusters of
individuals with the same opinion can be visualized as patches on a two-dimensional grid
Artificial societies
126
in which the nodes represent the members of the population. These clusters can be said to
have emerged and to form regularities.
In systems in which there is emergent behaviour, it is convenient to think of the
emergent properties of the system as influencing the actions of the agents. Thus, not only
do the agents actions at the local level, when aggregated and observed at the global
level, constitute the emergent behaviour, but also the global emergent behaviour can also
be said to influence the local actions of the agents, in a form of feedback. For example, in
Nowak and Latans simulation, one can speak of an opinion cluster influencing the
opinions of individuals outside the cluster. But it must be remembered that in the
standard complex system, there is no explicit mechanism for such feedback to take place:
agents are affected only by their local neighbours. The idea of feedback, or feed-down,
from global to local level is merely a convenient shorthand.
Levels of emergence
So far, for simplicity, it has been assumed that there are clearly distinct levels: the
micro and the macro. In fact, the situation is more complex than this. For example, it may
be the case that individual identity is best regarded as an emergent phenomenon, where
the micro-level agents are sub-cognitive, such as neurons. And societies are perhaps
best considered as emergent phenomena arising from the interaction of social institutions.
Thus it is better to consider a complex hierarchy of levels of emergence, rather than a
straightforward division between micro and macro.
A continuing worry of those who are interested in emergent systems, and particularly
of those who are simulating social process, is whether the emergent behaviours they
observe are in some sense programmed in and an inevitable consequence of the way in
which agents are constructed. Some criterion is required which will distinguish emergent
behaviour from behaviour which is predictable from the individual characteristics of the
agents. The description of complex systems suggests that a candidate criterion is that it
should not be possible to derive analytically the global emergent behaviour solely from
consideration of the properties of agents. In other words, emergent behaviour is that
which cannot be predicted from knowledge of the properties of the agents, except as a
result of simulation. This kind of definition, of course, provides an obvious explanation
for the popularity of simulation as an investigative technique in the field of complex
adaptive systems.
This criterion, however, is not sufficient to end the worries of those who are concerned
about whether they have really achieved emergence. For it is always possible that at
some future time analytical methods will be developed that can be used to derive global
properties from local ones. In a few cases, this seems to have happened. For example,
Forrest and Miller (1990) define a mapping between classifier systems (i.e. systems
consisting of nodes that use the genetic algorithm to synthezise patterns of connections
that collectively co-evolve to perform some function) and Boolean networks whose
properties are, relative to classifier systems, much better understood. This mapping opens
the way to predicting the emergent properties of classifier systems where previously
predictions were impossible. The example shows that if we define emergence in terms of
an inability to find an analytical solution, any particular emergent property stands the risk
of being demoted from the status of emergence at some time in the future. This suggests
127
that emergence may be neither a stable nor an especially interesting property of complex
systems: what are interesting are the systems macro properties and the relationship of
those macro properties to the micro ones.
Artificial societies
128
129
for example, the size of its immediate following. Some followers may
themselves be leaders. An agent may well appear in its own social model
as a leader or follower in particular groups. Also held within the social
model are beliefs about territories. There is no expectation that the
information held in an agents social model will be either complete or
accurate.
The social model is important because it has an effect on how an agent behaves in
relation to other agents. For example, an agents membership of a semi-permanent group
depends on the agent becoming aware of its membership of the group and then treating
fellow members differently from non-members.
Agents start without any knowledge of groups or other agents, but with a need to
obtain a continuing supply of resources, some of which can only be secured with the
co-operation of other agents. If an agent receives a request to co-operate with the plans of
another agent, it will do so, or it will attempt to recruit agents to a plan of its own. Agents
that successfully recruit other agents become group leaders and the other agents become
followers. Such temporary groups continue to work together, with group membership
recorded in the members social models, unless other processes (such as the agents
movement out of the proximity of other members) intervene.
A number of observations can be made about this simulation. First, in terms of the
definition offered previously, it is a complex system, although not an adaptive one.
Secondly, group membership is an emergent property of the behaviour of the agents.
Thirdly, agents pursue their own interests to obtain resources, forming groups as and
when necessary through negotiation with other agents. These groups have no existence
other than their representation in agents social models.
In particular, the agents, even though their membership in a group is recorded in their
social model, have no awareness of the group as an entity in its own right with which
they can reason. Agents not in a group have no way of recognising that the group exists,
or who its members are. In short, the simulation permits agents to reason about microproperties that they perceive in their local environment, and from this reasoning and the
consequent action, macro-level properties emerge, but because the agents have no means
of perceiving these macro-properties, the model as a whole fails to match the capabilities
of human societies.
Structuration theory begins to suggest what additional functionality could be built into
the agents. It proposes that actions are sedimented into structuring properties, such as
institutional practices, rules and typifications. Thus it would be necessary for the EOS
agents to recognise that they are members of groups and to realize what this implies
about members actions; and do so explicitly and in a way that results in a representation
they can subsequently reason with. Their own actions, as group members, would
reinforce (reproduce) these explicit representations of the group and its practices.
In the present simulation, a group emerges from agents actions, but it remains an
implicit property of the agents social models. As would be expected from a simulation
that adopts the individualistic approach implicitly, the existence of the group can be
observed by the researcher either by looking at the patterns of interaction at the global
level or by looking inside the heads of the agents, but it is not visible to the agents
themselves. An alternative way of implementing the simulation, rightly rejected by the
Artificial societies
130
research team, would have been to build the notion of a group into the system at a global
level so that the agents actions were constrained by their group membership. In this case,
which corresponds roughly to the holists position, the group would not have been an
emergent property of the simulation, but a constraining social fact.
A third possibility is to design the simulation so that agents themselves are able to
detect the emergence of groups and react (orientate) to this structure, so that they
become no less privileged in this respect than the observer. This latter course is the
approach that might be advocated by structuration theorists.
Conclusion
This chapter originated from a feeling that, despite the sophistication of the present-day
work, the models being used to simulate human societies are lacking in crucial aspects.
Few computational models of human societies have supported features which could be
argued to correspond to meaning, structure or action, although to sociologists,
concepts such as these are central to understanding human society. One of the
suggestions of this chapter is that more sophisticated models of human interaction are
essential if we are to progress with the simulation of human societies. Human societies
are not just ant hills writ large. However, pinning down clearly, and in a computationally
precise way, the nature of the difference between non-human and human societies is not
straightforward.
The second suggestion of this chapter is that sociological debates about the nature of
human society can contribute to our understanding of these issues, even if a certain
amount of effort is required to translate these debates into forms that are accessible to
computational modelling. We have seen that one theorystructurationprovides some
useful suggestions about how the relationship between macro- and micro-level properties
might be conceptualized and even some broad indications about how a simulation might
be implemented. It must be noted, however, that structuration theory is certainly not the
only position on these issues to enjoy respect among sociologists, and that other presentday ideas may be equally, or more, fruitful.
The third suggestion is that one capacity of humans which is probably not widely
shared by other animals is an ability to perceive, monitor and reason with the macroproperties of the society in which they live. It seems likely that this ability is related to,
and is perhaps even a consequence of, the ability of humans to use language. Much of the
lexicon of modern adults is, for example, composed of concepts about macro-properties.
The fourth and final suggestion, lurking behind the rest of the argument, is that human
societies are dynamic; that is, they are (re-)produced by people as they proceed through
their everyday life. Routines, typifications, structures, institutions and so on, are not
reflections of some external reality but are human constructions. In particular, the
macro-properties that are recognized by members of human societies are not fixed by
the nature of the agents (as they would be in other complex systems), but are an
accomplishment, the outcome of human action.
131
Acknowledgements
I am grateful to the EOS project team, especially Mike Palmer and Jim Doran, and the
participants at SimSoc 92 and 93 for opportunities to discuss these issues.
Notes
1. I am simplifying here; meanings are indexical and negotiated in the context of use in a way
which we do not yet properly understand, at least from a computational perspective. See
Hutchins and Hazlehurst, Chapter 9 in this volume.
Chapter 9
How to invent a lexicon: the development
of shared symbols in interaction
Edwin Hutchins and Brian Hazlehurst
Human language provides, among other things, a mechanism for distinguishing between
relevant objects in the natural environment. This mechanism is made up of two
componentsforms and meaningswhich must be shared by the community of
language users. The lexicon constitutes much of a languages form, but its description as
a set of shared, form-meaning resources for communication is seldom addressed by
formal models of language. This chapter presents a simulation model of how shared
symbols (form-meaning pairs) can emerge from the interactions of simple cognitive
agents in an artificial world. The problem of creating a lexicon from scratch is solved by
having cognitive agents that are capable of organizing themselves internally share their
expressions of visual experience in interaction through learning to classify a world of
visual phenomena. The model is seen as an instantiation of a theory of cognition which
takes symbols to be a product of inter- and intra-individual organizations of behaviour,
the result of cultural process.
Below, we first present a theoretical stance that has been derived (elsewhere) from
empirical investigation of human cognitive phenomena. This stance leads to the
articulation of some simple information-processing constraints which must hold for
lexicons and their users. We present later two simulations employing communities of
simple agents that allow us to model how a lexicon could come into existence or emerge
from the interactions between agents in an extremely simple world of experience. The
model allows us to explore issues which involve interaction between group-and
individual-level properties of cognition, social organization and communication.
Subsequently, we briefly review some findings from the literature on lexicons of natural
language and conclude the chapter with an evaluation of the model as an instantiation of
the theoretical stance.
133
situations (Byrne & Whiten 1988, Levinson in press). This kind of relationship between
the social and the cognitive is important, but it is still centred on the notion of the
individual as the primary unit of cognitive analysis. The social world is taken to be a set
of circumstances outside the individual, about which the individual reasons. What we
intend, instead, is to put the social and the cognitive on an equal theoretical footing by
taking a community of minds as our unit of analysis.
This new perspective permits two things to happen that are not possible from the
traditional perspective. First, it permits inquiry about the role of social organization in the
cognitive architecture of the system, and allows description of the cognitive
consequences of social organization at the level of the community (Hazlehurst 1991,
Hutchins 1991). Second, it permits symbolic phenomena that are outside the individuals
to be treated as real components of the cognitive unit of analysis (Sperber 1985,
Hazlehurst 1994, Hutchins in press).
Making this move also presents an opportunity to view language in a new way.
Cognitive science generally takes the existence of language as a given and concerns itself
with the sorts of cognitive processes that must be involved when an individual processes
language (in production and comprehension). From the perspective of the community of
minds as cognitive system, a languages information-bearing capacity, conventions of
use, functionality, distribution, variability, etc., all become determinants of the cognitive
properties of the community because these things are instrumental in determining where,
when, and what kinds of information move through the system (cf. Freyd 1983). This
attention to the movement of information in the larger system necessarily brings the
material world back into play, since, having acknowledged symbols outside the head, we
now must take seriously their material nature. Furthermore, we need not, indeed must
not, ignore the means individuals have or develop for incorpo-rating symbols into private
as well as collective organizations of behaviour. By redefining our unit of cognitive
analysis, it seems to us that progress may be made towards reuniting the social and the
material with the cognitive (cf. Goody 1977).
Artificial societies
134
can arise where none existed before. These structures are products of a system whose
organizing dynamics are an interplay between intra-individual and inter-individual coordinations.
In the presentation, these structures will be referred to as terms, descriptions, words, or
patterns of acoustic features. There is, however, no strong a priori commitment to any
particular level of linguistic representation, and the structures described might just as well
be thought of as patterns of denotational or even relational features. Each of the above
terms takes its meaning from a claim about the function of these public representations in
this artificial world. We take this stance to be an important methodological and
theoretical component of this work. Representations do not get to be symbols by virtue of
us creating them or calling them such, but rather, by our reading of what they do for the
members of a community of artificial cognitive agents (Clancey 1989).
The model is based on six central theoretical assumptions, derived from a theory of
distributed cognition (Hazlehurst 1991, 1994; Hutchins 1990, 1991, 1993, in press;
Hutchins & Hazlehurst 1991; Hutchins & Klausen in press):
1. No mind can influence another except via mediating structure. (The no telepathy
assumption.)
2. No social mind can become appropriately organized except via interaction with the
products of the organization of other minds, and the shared physical environment.
(The cultural grounding of intelligence assumption.)
3. The nature of mental representations cannot simply be assumed, they must be
explained. (The shallow symbols assumptionin contrast to the deep symbols
assumption which brings symbols into the language of thought as an article of faith
rather than as a consequence of cultural process.)
4. Symbols always have both a material and an ideal component. The material component
is what makes form, structure and differences possible. The ideal component is a
function of the stance that organized individuals take toward these material forms.
(The material symbols assumption.)
5. Cognition can be described as the propagation of representational state across
representational media that may be internal to or external to individual minds. (The
distributed information-processing assumption.)
6. The processes that account for the normal operation of the cognitive system should
also account for its development through time. (The no developmental magic
assumption.)
Below we present a computer simulation that is an implementation of these assumptions.
It turns out to be a very robust procedure by which a community of individuals can
develop a shared set of symbols. The simulation is not to be taken seriously as a model of
any part of human history. It is the simplest possible scheme that captures the essential
properties of the system being modelled. The simulations are not to be taken as
representations of actual human cognition, language, culture or experience, but as
existence proofs that a particular kind of process is capable of producing a particular sort
of outcome, in this case, a community with a shared lexicon. One is certainly free to
question the extent to which it is reasonable to map this process on to plausible courses of
human behaviour and in the discussion section we consider ways in which the model is
consistent with observations about human language.
135
Artificial societies
136
dividual and those in another individual (as is the case in communication), or between
one set of internal structures in an individual and another set of internal structures in that
same individual (as is the case in using written records as a memory aid, for example).
Internal structures provide bridges both between successive artefactual structures and
between natural and artefactual structures. Following Sperber (1985:76) we may say that
A representation [artefactual or internal structure] is of something [natural, artefactual,
or internal structure] for some information processing device [internal or artefactual
structure].
Connectionist networks as simple agents
During the past eight years, developments in computational modelling employing
connectionist networks have made it possible to think in new ways about the relationships
between structure inside a system and structure outside.4 A network is composed of two
kinds of mathematical objects: units and connections. Units take on activation values,
generally in the range of 0.0 to 1.0, and pass activation along one-way connections to
other units. This passing of activation along a connection is modulated by a real value
associated with that connection, the connection strength or weight. The passing of
activation from input units to output units of the network can be treated as the
implementation of a function, and viewed as the networks behaviour in a given
environment. Through modification of the networks weights, in time, the network adapts
to the shape of that environment.
Connectionist networks of a class called autoassociators have particularly nice
properties with respect to the problem of discovering and encoding structural regularities
in their environment. Autoassociator networks learn to duplicate on the output layer the
identical pattern of activation pre-
137
sented to the input layer. Figure 9.2 shows a simple autoassociator network. It consists of
three layers of units.
Input units are on the left, output units on the right, and hidden units in the middle.
Targets are real valued vectors which are structurally similar to the output and input
layers but, like inputs, are thought of as information external to the network. Targets are
part of the environment which the network is made to learn (see below)they provide
the teaching signal. In the case of autoassociators, the targets are simply the input patterns
themselves, allowing the network to learn about the environment with no additional
teaching signal.
Limitations on space make a full description of this kind of information processing
system impossible. The following sentences will, we hope, convey the style of
computation entailed, if not the details.
Every unit in the input layer of a network has a unique connection to every unit in the
hidden layer, and every unit in the hidden layer has a unique connection to every unit in
the output layer (see Figure 9.2). The strengths of these connections can be adjusted. The
activations of the input layer are set by external phenomena. The activations of the other
units are determined by the activations of the units from which they have connections and
on the strengths of those connections. Starting from random connection strengths, the
task for the network is to discover a pattern of connection strengths that will produce the
desired output (i.e. the target) in response to a given set of inputs. Incremental
improvement in accomplish-ing this task is referred to as learning. The networks
modelled here use a procedure called the back-propagation of error to find an
appropriate set of connection strengths. In this scheme, the output produced is compared
to the target,5 and the difference between output and target is an error in the networks
ability to perform this input-output mapping. The connections are then adjusted to reduce
this error on future trials at this task. The problem for the network can be viewed as one
of finding a set of weights which meets simultaneously the constraints imposed by all of
the input-output mappings it is made to perform.
Meeting the first constraint: words must discriminate between objects
in the environment
Rumelhart, Hinton & Williams (1986) have shown that under certain conditions, the
activations of the hidden layer units of autoassociator networks converge on efficient
encodings of the structural regularities of the input data set. That is, the connections
between input and hidden units must produce activations at the hidden layer which can be
used by the connections between hidden and output units to produce the target, under the
constraints of the function that propagates activation. For example, given any four
orthogonal input patterns and an autoassociator network with two hidden units, the
hidden unit activations for the four input cases should converge on {(0, 0) (0, 1) (1, 0) (1,
1)}. This is because the network must use the activations of the hidden units to encode
the four cases and the encoding scheme attempts to distinguish (optimally) among the
cases.
Producing these efficient encodings is equivalent to feature extraction. That is, what
the networks learn is how to classify the input data in terms of distinctive features or
principal components. If we fold an autoassociator in half, and allow the hidden layer to
Artificial societies
138
139
other to be the target, then consensus will result. This satisfies the constraints that A1=B1,
A2=B2,, Am=Bm.
Artificial societies
140
for comparison with As verbal representationis also being used to produce a visual
output by feeding activation forward to the visual output layer. The effects of this
learning on Bs future behaviour can be stated as: (a) in this context it produces a
representation at verbal output more like what was said by A, and (b) produces a
representation at visual output more like the scene itself.7
Over time, by choosing interactants and scenes randomly, every individual has the
opportunity to interact with all the others in both speaking and listening roles in all visual
contexts. The effect to be achieved is for the population to converge on a shared set of
patterns of activation on the verbal output units that makes distinctions among the m
scenes. That is, we hope to see the development of a consensus on a set of distinctions.
Below we discuss the results of two different sets of simulations. Simulation 1 was
based on a more complex network architecture and set of scenes. Simulation 2 used the
simpler network architecture already shown in Figures 9.3 and 9.4. The scenes of
Simulation 2 were the four orthogonal vectors (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0) and (0,
0, 0, 1). The purpose of discussing Simulation 1 is to demonstrate the qualitative effects
of evolving a lexicon within a (relatively) complex system space. The purpose of
discussing Simulation 2 is to explore more analytically the nature of this kind of dynamic
system in a more manageable system space.
141
Artificial societies
142
143
Artificial societies
144
145
representations (i.e. it measures the average term distinctiveness, across all scenes as
denoted by each individual, averaged across all individuals). Avg2 is a measure of the
variability in the community of individuals verbal representations (i.e. it measures the
average term variability for each scene across individuals, averaged across all scenes).12
Our expectation is that Avg1 will tend toward 1.0 (the maximum difference in terms used
across scenes), and Avg2 will tend toward
Artificial societies
146
activation versus Unit 2 activation) for each of the four scenes in the world. Each
trajectory on a graph represents one individuals term for that scene, parameterized by
time. The trajectories all begin near the middle of each graph (verbal output activations
near [.5, .5]) because, at the beginning of the simulation run, individuals are responding
to the scenes with unorganized connection weights.13 As time proceeds, the trajectories of
each graph head for one of the four corners (i.e. consensus regarding each scene
increases). Furthermore, each graphs trajectories must reach a corner unoccupied by the
trajectories of the other three graphs (i.e. term similarity decreases). Notice how two of
the emerging lexicons terms (Figures 9.9(a) and 9.9(b)) compete with each other for a
place in the (1, 0) corner before the term representing scene (0, 1, 0, 0) of Figure 9.9(b)
finally wins this competition.
147
again, the networks are responding roughly the same (and uninformatively) at the
beginning of the simulation. Between 1,000 and 2,000 interactions, as good
representations for discriminating between scenes emerge (Avg1, the mean variability in
each individuals descriptions for different scenes, goes up), the degree of consensus goes
down (Avg2, the mean variability in descriptions representing the same scene, across
individuals, goes up). As the rate of learning to discriminate between scenes slows down
(Avg1 approaches the asymptote), the representations which serve as individuals verbal
targets change more slowlythey become easier targets to follow, and consensus begins
to emerge (Avg2 begins to descend again).
The effect of varying community size
We are now in a position to ask how the dynamic system depends on population size n,
given that we have fixed all the other system parameters
148
Artificial societies
149
as discussed above. Figure 9.11 shows the results of three different experiments, each
entailing a sample of 15 simulations.
The simulations of each experiment were run with community member sizes n of 5, 10
and 15, respectively. The means and one standard deviation error bars for the 15
observations of each experiment (sampled at regular intervals within each simulation) are
shown for the two measures of lexicon structure, Avg1 and Avg2. In all three
experiments, individuals have participated (on average) in the same number of
interactions (namely, 2,000) by the end of the time frame shown in Figure 9.11. The
general pattern is the same as that seen in the single simulation run of Figure 9.10. Of
particular interest is the nature of the decay in the lexicon formation process shown by
changes in the two measures of lexicon structure, Avg1 and Avg2, as community size
gets larger. This decay is caused by the greater difficulty of organizing large communities
rather than small ones. Each experiment displayed in Figure 9.11 shows that Avg1 and
Avg2 of the average community (represented by the plotted mean values of the 15
simulations in each graph) vary smoothly and asymptotically as a function of the number
of interactions. Therefore, the final steady-state of each experiment can be approximated
reasonably well by the final mean value of each graph in Figure 9.11. Taking the three
points so collected for each measure (Avg1 and Avg2), it appears that the decay in the
ability of a community to form
Artificial societies
150
these simulations do not address. (But see Hutchins & Hazlehurst 1991, for how such a
lexicon may be embedded in a larger context of community use.)
Discussion
The need for a critical period of language learning
We have experimented with adding new individuals with random weight-structures to
communities that have already developed a lexicon. Depending on the size of the
community, the addition of the new individual may have quite different effects. The
behaviour of a new individual added to a large community with a highly shared lexicon
will be entrained by the behaviours of the other members of the community and the
newcomer will learn the shared lexicon. A new individual added to a small community
may completely destroy the previously achieved solution. After such an event the
community may or may not be able to recreate a well-formed lexicon with the new
individual.
In running these simulations we found ourselves wishing that there was some
principled way to reduce the learning rate once individuals had learned the language. In
particular, one would like to reduce the learning rate at the point in the life cycle where
individuals are likely to encounter many interactions with disorganized individuals.
This would amount to implementing a critical period for language learning so that
individuals learn less from the linguistic behaviour of others once they have reached
sexual maturity.14 Perhaps evolution has engineered something like this into our species.
We did not implement a critical period, however, because to do so seemed arbitrary and a
violation of one of the core premises: the processes that account for normal operation of
the system should also account for its development through time. In more complex
situations, such as that of biological evolution, where adaptive searches are conducted in
parallel at many levels of specification, it may be reasonable to expect violations of this
premise.
The formation of dialects
Occasionally, communities fail to create a well-formed lexicon. This happens because
sometimes the random initial starting points of the networks in a community are
incompatible with each other, and unlucky choices of the interaction protocol lead to
divergence in the verbal representations of these individuals, which cannot be
overcome15. In this case, the kind of language that one individual is predisposed to learn
is not the sort that the other is predisposed to learn, and given their starting points and
learning experiences, they never find a solution that can be shared.
The fact that some pairs of initial weight configurations are more compatible than
others suggested that it might be advantageous to let the individuals discover and seek out
those with whom they are compatible (as demonstrated in the similarity of their
behaviours, i.e. their verbal representations). A sequence of experiments was run in
which the choice of a listener for each speaker was biased in favour of those who had a
history of speaking (i.e. using descriptions in former interactions between the two) which
151
was similar to the speakers own history of utterances. The result of implementing this
interaction protocol was the formation of dialects, or clusters of individuals within which
there is consensus on the descriptions and their referents. Since individuals interact more
with like minded others, they are more susceptible to learning from the classification
behaviours of those who act as they themselves do. We take this to be a very primitive
implementation of what Sperber (1985) called ecological patterns of psychological
phenomena.
The acquisition of a coherent lexicon
Consistent with the observations reported above about dialect formation and the
education of new individuals, is an interesting interpretation of the models performance
regarding the ontogenetic problem of learning form-meaning pairs that are already
established in the world. In particular, it seems that the complement to the need for a
critical period of language learning (which ensures that the experience of older
individuals is not lost by attending to the unorganized ramblings of novices) is the need
for a set of consistent models during language acquisition (which ensures that novices are
exposed to a system that is coherent, and therefore learnable).
Creating a lexicon from total lack of organization (demonstrated in Simulations 1 and
2), seems to be easier than creating a functional lexicon from a system that is organized
in the wrong way. That is, novices introduced into the system who must accommodate a
wide range of organized variance (as with dialects) have a hard time learning any part of
the form-meaning system. On the other hand, new individuals exposed to well-formed
lexicons (or one well-formed dialect) gain significant leverage on the problem of learning
the form-meaning system via the mediation of consistent and well-formed terms. In fact,
experiments were run which showed that the strictly visual problem of classifying scenes
in the simulation world is simplified by the use of consistent and well-formed terms
individuals abilities to learn the visual classification problem are enhanced by the
existence of a coherent lexicon.
This fact of the simulation seems to be because of the nature of the decomposition of
the principal component being conducted by the autoassociator in its learning (Chauvin
1988). By grounding this process in the constraints imposed on the hidden layer by
coherently organized targets, the decomposition process is accelerated significantly There
is a sense, then, in which the structure of the lexicon is an important vehicle for the
learning of the visual classification problem.
In the real world, we can cite two kinds of evidence that are consistent with this aspect
of our models behaviour. The first has to do with the nature of language-acquisition
environments. Children generally do learn language, over an extended period of time, in a
relatively homogeneous population of language users. This is a fact which follows from
the general nature of the social organization of our species. Second, there appears to be a
good deal of empirical evidence for the fact that children must (or at least that they act as
though they must) accommodate each and every form-meaning pair that they encounter
(Clark 1983, 1987; Slobin 1985). That is, the language-acquisition process apparently
requires the learner to assume that words contrast in meaning, and children use this as a
resource in classifying the objects or events which words denote (Clark 1987).
Artificial societies
152
Clark (1987) investigated the natural-language version of our notion of a lexicon being a
shared set of distinctions. She provides empirical evidence consistent with the claim that
natural-language lexicons exhibit two pervasive features:
1. Every possible pair of forms contrast in meaning (the principle of contrast).
2. For certain meanings, there is a conventional form that speakers expect to be used in
the language community (the principle of conventionality) (Glark l987:2).
The general claims made are that the systems which generate human lexicons are
efficient (all words contrast in meaning) and conservative (established words have
priority, and innovative words fill lexical gaps as opposed to replacing established words
with the identical meanings). Evidence for these principles is cited from a wide range of
natural-language phenomena, including the observations that:
1. Lexical domains emphasize semantic contrasts.16
2. Syntactic constructions create semantic contrast; differences in form mark differences
in meaning at both the lexical and the syntactic levels (Clark 1987:6).
3. Well-established irregular forms in a language are maintained in the face of many
resources (paradigms or patterns) for simplifying the language through regularization.
4. Innovative (new) words emerge as a consequence of a failure in an existing lexicons
capacity for conveying the appropriate meaning (i.e. due to an inability of the existing
words to establish the appropriate contrasts).
These principles of human language have parallel entailments for the acquisition of
language:
1. Children rely on contrasting terms to tune their understanding of semantic fields to
adult levels of specificity and generality.
2. Children assume (or at least act as though they are assuming) that new terms contrast
with those that they already know.
3. Children reject (both across and within languages) terms which they understand to be
synonyms with terms that they already know.
4. Children productively generate novel terms to fill expressive needs, but these terms
converge toward conventional usage as their language development proceeds.
We find Clarks analysis to be in general agreement with the phenomena modelled and
suggested by our simulations, although her observations clearly exceed the range of
phenomena covered by our simple simulations.
Grounding meaning in communicatory practices of the community
We find Clarks analysis lacking in two respects: the first is made explicit by our model;
the second is suggested by the general theoretical framework we have adopted, although
not specifically captured in the simulations we have presented here.
First, Clarks formulation tends to equate language use with form-meaning pairing,
which lapses into a notion of meaning that is structurally derived rather than grounded in
communicatory practice. This appears to stem from an analysis which takes language
153
Artificial societies
154
155
capture well the behaviour of this system. The structure of the natural world is fixed in
this model, but the internal structures and the artefactual structures co-determine each
other and co-evolve in the development of the lexicon. In the broadest sense, the solution
arrived at was determined by the structure of the natural world as manifested in the
phenomena encountered, in the random initial configurations of the internal states of the
individuals, and in the instantiation of who actually learns from whom in the community.
The process of developing a lexicon in this model is a process of propagating transformed
representations of naturally occurring structure throughout a system that also contains
artificial structure.
Finally, even when two networks are in complete agreement with one another about
the use of the lexicon, each has a unique internal structure. Individuals in our model are
able to use the lexicon without needing to share the internal structures that enable that
use. Through learning from each other in interaction, individuals become functional
equivalents, not structural replicates of each other. There is no need to posit the existence
of grandmother neurons which are responsible for the like behaviours of autonomous
individuals. On the other hand, behaviour is shaped by the constraints of autonomous
individuals interacting in (problematically) shared environments.
Within this theoretical framework, we claim that meaning can be re- trieved from the
unattractive positions of being equated with:
(a) the results of a (usually innate) private language of mental symbols that stand for an
uncontested, fixed and non-social objective reality (Fodor 1976; cf. Lakoff 1987); or
(b) an unproblematically shared, static (i.e. a non-developmental, non-historical, and
often non-social) semantic knowledge base (ethno-science and much of cognitive
anthropology; cf. Hutchins 1980); or
(c) a strictly public (i.e. superindividual, non-mental) symbol system (Geertz 1973; cf.
Shore 1991).
Meaning in our model is an evolving property of the interaction of internal, artificial and
natural structures. At any point in the evolution of such a system we take meanings to be
characterizations of the functional properties of individuals is--is the environment
(including each other). Meaning, for each individual, describes a range of possibilities for
action. In a more complicated world, possibilities would be constrained by representation
of the consequences of these actions, something we understand to be true of human
meaning systems. In the limited world we have created, the range of possibilities for
action is restricted to producing something like monolexemic utterances in the contexts of
shared binary scenes. In the real world the possibilities are much greater, some of which
we hope to address in future work.
Appendix
In a community of n agents who live in a world with m scenes, the communitys language
at time t (the recording of all agents descriptions of the m scenes) can be written in
matrix form as:
Artificial societies
156
where Si,j(t) is agent js description of the scene i at time t. Then the sequence of matrices,
{L0, L2,, L} represents the evolution of the communitys language from time t=0 to t=.
At any time t, the language Lt can be analyzed for at least two properties:
(a) Avg1, the average difference between all pairs of descriptions of scenes for each
individual; and
(b) Avg2, the average difference between pairs of individuals descriptions of the same
scene, for all individuals and scenes.
Avg1 gives a measure of individuals abilities to use terms to distinguish between scenes,
while Avg2 gives a measure of the communitys ability to agree on the terms used to
identify scenes.
More formally, assume Si,j(t) and Sl,p(t) are real valued vectors of length , and define a
distance metric:
where
is the kth real value in vector Si,j (t).
Then,
where P2(m) is the set of all pairs of integers from 1 to m, and (m2m)/2is the size of the
set. Similarly,
157
Acknowledgements
Research support was provided by grant NCC 2591 to Donald Norman and Edwin
Hutchins from the Ames Research Center of the National Aeronautics & Space
Administration in the Aviation Safety/Automation Program. Everett Palmer served as
technical monitor. Additional support was provided by a fellowship from the John D. and
Catherine T.MacArthur Foundation.
Notes
1. In fact, a description is a vector of real values which we imagine to be articulatory features
capable of generating a monolexemic word, a piece of agent-created structure, in this
artificial world. The act of referring is a property of our construction of the worldwe
have built into the world the need for agents to internalize visual experiences and code that
organization in externally realizable structures. This, of course, bypasses a long history of
evolution which might select some of these properties as internal properties of agents.
However, in granting these (on face value, plausible) assumptions we are better able to
address cognitive and cultural issues with our simulations.
2. This model does not deal with homonyms or synonyms. However, note that it is a matter of
current debate whether such categories in fact exist in human language when the perspective
of language use and learning is taken to examine instances claimed to be synonyms or
homonyms (Clark 1987, Slobin 1985).
3. See Appendix for a more formal treatment of this analysis.
4. The best background work on connectionism is by Rumelhart, McClelland et al. (1986), and
McGlelland et al. (1986). The behaviour of autoassociator networks is thoroughly analyzed
in Chauvin (1988).
5. For an autoassociator, the target is identical to the input, thus reducing the problem to an
identity mapping on the input set.
6. We thank Elizabeth Bates (personal communication, February 1991) for coining the term
public hidden units for this construction.
7. Implementing this combination of error signals is straightforward. One error signal is
computed from the difference between produced word and target word, the other error signal
is the usual error back-propagated from the visual output layer. These two signals are simply
added together and back-propagated to the visual input layer.
8. It is a well-known result from connectionist history that two layers of weights are required to
perform mappings from input to output which are not linearly separable (Rumelhart et al.
1986). The range of verbal representations that individuals are attempting to map to in this
simulation may constitute such a set, thus requiring the extra hidden layer to perform
properly. We say may, because the mapping itself is evolving as the lexicon is being
constructed. Another reason for the required extra layer has to do with the large compression
Artificial societies
158
of data from input (36 units) to verbal input/output layer (4 units). This compression tends to
swamp the verbal output layer with large activation values (even for an untrained network),
reducing the networks flexibility to learn. Since the learning problem is composed from the
behaviours of other (now inflexible) agents, the community is unable to converge upon a
shared lexicon.
9. The consensus in the starting state of the simulation is a product of the fact that random
weights in the network tend to produce mid-range output values regardless of input to the
network.
10. See The formation of dialects, section (below) for a discussion of some observed
exceptions.
11. The learning rate is a parameter controlling the magnitude of the effect that one learning
trial has on a network, i.e. the scale of magnitude by which changes are made to weights on
each learning trial. The learning momentum is a parameter which influences the effects that
variability in the learning set has upon network learning performance. That is, the learning
momentum parameter determines the scale of magnitude by which recent learning trials
continue to affect the current learning trial. (See McClelland (1988), for implementation
details regarding these learning parameters.) These two parameters could, conceivably, vary
among individuals, perhaps also as functions of time. In the simulations of this paper we
chose to fix these parameters, not letting them vary across individuals or time, in order to
simplify the task of understanding the set of dynamic systems we are dealing with.
12. See Appendix for a more formal definition of these measures.
13. There has been no learning yet, and all individuals begin with random weight assignments to
their connections, therefore all units respond at mid-range levels of activation.
14. Of course, an alternative solution might be simply for individuals not to focus their learning
lens too narrowly on any other single individual (or homogeneous group of individuals). The
problem with this solution is that it works for organized (adult) individuals but does not
work for disorganized (novice) individuals. See The acquisition of a coherent lexicon
section.
15. The frequency of this phenomenon is dependent upon many parameters of the simulation
which determine the learning trajectories of the networks (see Simulation Two, above, for a
listing of these parameters). Included here are the learning rate and learning momentum
parameters of the network learning algorithm. The stability of the environment being learned
also plays a critical role in determining the probability of networks becoming stuck in their
search for a set of weights to perform the tasks required.
16. Clark claims that true synonyms do not exist in natural languages.
17. This fact seems to get Clark into trouble in her critique of Slobins notion of
unifunctionalitywhich denies the existence of multiple forms carrying the same meaning
(something Clark agrees with), and denies the existence of multiple meanings being carried
by the same form (something Clark disagrees with; 1987:25). Clark claims, for instance, that
the English inflection -s is a form used productively to map on to the concepts of plurality
and possession. Clarks argument here is based solely on evidence from the structural
regularities of parts of English morphology, effectively ignoring the possible communicative
and learning functions which act to create contrasts even here. These are the properties of
natural language that Clark relies upon to build her own case elsewhere in the paper.
Furthermore, Clark (1987:26) utilizes a logical analysis of semantic feature inheritance as an
argument for rejecting Slobins denial of true homonymy in natural language. This seems to
be inconsistent with her use of languages communicative functions found elsewhere in her
paper.
18. Work on the slightly longer step of developing propositional representations and letting
symbols break free of the phenomena to which they refer is now in progress (cf. Hutchins &
Hazlehurst 1991).
159
19. Here we must acknowledge a hedge on our own claims. The model, as it stands, collapses
semantic and phonological representations. There is no distinction between what an agent
conceives of the scene and what it says about the scene. Similarly, what an agent says is
unproblematically heard by the other agent participating in the interaction. This is, strictly
speaking, a violation of the no telepathy assumption, and the separation of internal and
artifactual structure. Having acknowledged this discrepancy we can state that the model, as it
stands, gains no explanatory power from this conflation and is therefore not a violation in
principle. We have experimented with architectures that produce phonological
representations from semantic representations, and vice versa. Implementation of these
constructions do not render invalid the general claims of the model presented here.
Chapter 10
MANTA: new experimental results on the
emergence of (artificial) ant societies
Alexis Drogoul, Bruno Corbara, Steffen Lalande
The MANTA project, already described in Drogoul et al. (1992a, b) and Ferber and
Drogoul (1992), is an application of the EthoModelling Framework (EMF) to the
simulation of the social organization of ant colonies. EMF is based on the principles of
multi-agent simulation (for a formal definition, see Ferber & Drogoul 1992), which
means that each individual in a population is represented by an artificial entity whose
behaviour is programmed (computationally speaking) with all the required details (for
other works on multi-agent simulation see Gilbert & Doran 1994, Hogeweg & Hesper
1985, Collins & Jefferson 1991). Multi-agent simulations are of most help in modelling
situations in which individuals have different complex behaviours and where the
interactions involve so much non-linearity that they cannot be described easily within a
mathematical framework. Our aim with MANTA is to test hypotheses about the way in
which social structures, such as a division of labour, emerge as a consequence of the
behaviour and interactions of individuals. In other words, we want to evaluate the
minimal set of causes that has to be provided at the micro-level to observe definite
structures at the macro-level.
Some early simulation experiments conducted with a preliminary version of MANTA
have been presented in Drogoul et al. (1992a, b). They showed that it is possible to obtain
the emergence of a division of labour within a nest of simple ants, i.e. ants provided
with only three tasks. Nevertheless, our goal at that time was not to simulate the
complexity of the social organization as it is observed in real nests but to demonstrate the
relevance of our approach. This chapter presents the preliminary results of a more
ambitious set of simulation experiments on the mechanisms of sociogenesis (Wilson
1985) and division of labour. These experiments have been conducted using a new
version of MANTA, in which the ants are more realistically represented than in the
previous one, being provided with a larger set of behaviours.
MANTA
161
engage in more than one task at a time. The tasks are triggered by stimuli that can have
variable strengths and may be either internal or provided by the environment. Stimuli are
associated with pre-existing motivations, expressed in terms of tasks weights, thresholds
and activity levels.
The weight stands for the relative importance of the task within the agent. In addition
to being a component of the basic behaviour reinforcement process1 provided for the
simulated ants, the weight is also used to accumulate the agents previous experiences of
a given task. A high weight indicates that the agent is specialized in the task. The weight
is used to compute the activation level of a task when it is triggered by a stimulus. Its
reinforcement is considered to be a form of long-term positive feedback.
The threshold is viewed as an indicator of the motivation2 of the agent to perform a
given task. This motivation is increased continuously as long as the task is not activated,
and decreased whenever it becomes active. As a result, the threshold of the neglected
tasks decreases and the threshold of the current task increases during the task selection
process. These operations are considered to be short-term positive and negative feedback.
Finally, the activity level is viewed as an indicator of the agents motivation to
continue the task it is performing. The activity level is initialized with the activation level
of the task and then decreased continuously as long as the task remains active. When it
reaches zero, the task is stopped. Hence a task is considered to be capable of being
activated when its threshold is lower than its weight multiplied by the strength of its
triggering stimulus (that is, its activation level). And it is selected when this activation
level surpasses the activity level of the current task. The agent then switches from the
previous task to the new one.
Communication
As defined in the kernel of EMF, agents do not communicate with one another. They just
drop stimuli into the environment, and these stimuli may or may not trigger the
behaviours of other agents. Dropping a stimulus results in the creation of a gradient field
around the emitter (by diffusing this stimulus in concentric circles around it), and in a
way analogous to the chemical, visual or oral propagation of information. Some of these
communications are part of the agents behaviour (they can deliberately choose to
propagate stimulus to attract or repulse other agents), and some of them are independent
of their behaviour (for example, an agent will always be visible to other agents, be they
friends or enemies, if they can see it). The propagation of stimuli is only stopped by
obstacles such as walls.
Implementation
EMF follows an object-orientated approach, although the agents do not use messagepassing as a means of communication. The agent objects are instances of classes that
inherit from Ethobehaviour or one of its subclasses which implement the abstract model
of behaviour they are going to use (cf. the use of objects described in Chapter 5). Each of
these classes represents a species of agent and provides the agent with information about
primitives, tasks, stimuli and other domain-dependent pieces of knowledge shared by all
agents. This information can either be described in the agents class, or inherited from
Artificial societies
162
higher classes. The subclasses of Ethobehaviour are divided into two sets: abstract
classes and concrete classes. Abstract classes (such as Ethobehaviour) cannot be
instantiated. They are used to define the primitives and knowledge associated with
primitives. Concrete classes, which can have instances, inherit these primitives and use
them to define the tasks (behaviours) of their agents. An example of such a hierarchy is
provided in Figure 10.1, in which all the classes whose names end with Behaviour are
abstract classes.
Interfacebehaviour implements the protocols of the agents user-interface capacities
(graphical trace, inspection, etc.). Locatedbehaviour provides all its subinstances with the
ability to be in an environment and to act on it (propagating stimuli, for example).
Curingbehaviour implements the primitives needed by agents that will have to cure other
agents or receive care from them. Feedingbehaviour implements primitives needed by
agents that will have to feed themselves, feed another agent or be fed by one. Maturing-
MANTA
163
behaviour provides its subinstances with the capacity to grow old (and subsequently die!)
and to become another agent. The notion of time is implemented in this abstract subclass,
as an internal stimulus. The subclasses of these last three classes must define some
domain-dependent pieces of knowledge, such as the average expectation of life of their
instances, their needs for food and so on. This knowledge will be used to compute
implicit stimuli that will always be propagated by these agents. These stimuli are named
after the type of agent (ant, larva, etc.) prefixed by cure-, hungry- or maturing-.
Movingbehaviour gives agents the possibility of moving in their environment.
Sensingbehaviour implements two primitives for following or fleeing a gradient field.
Carryingbehaviour implements the primitives needed for carrying other agents.
The MANTA classes
The concrete subclasses can be divided into three sets. FoodAgent, DeadAntAgent,
HumidityAgent and LightAgent represent environmental agents, that is, agents that do not
perform any behaviour other than that of propagating their stimuli (#deadAnt, #food,
#light and #humidity). These stimuli can trigger some ants behaviour (#food) or be used
as guides for behaviour
Artificial societies
164
MANTA
165
triggered by an internal stimulus whose strength is a periodic function of the time (the
task should be triggered every six hours).
Sociogenesis
Definition and assumptions
In the foundation process or sociogenesis (a word coined by Wilson (1985) by analogy
with morphogenesis), the newly-fertilized queen alone initiates a new society. In
Ectatomma ruidum, a tropical ant belonging to the phylogenetically primitive subfamily
of the Ponerinae and whose societies are the natural counterparts of MANTAs simulated
ones, the foundation is said to be semi-claustral. The founder leaves her nest to provide
the food that is necessary for herself and, above all, for the larvae (in a claustral
foundation the queen provides food to the larvae by using her own internal reserves
derived from the degeneration of her wings musculature). Furthermore, as we have
shown elsewhere (Corbara 1991), the queen generally continues to forage after the
emergence of the first workers. Sociogenesis is then a remarkable challenge for the
queen: she has to face a very complex situation alone during the first stages of the
society, when she has to take care of the whole brood as well as going to find food. And
it is not really surprising to see that, even in laboratory conditions, only 15 per cent of
sociogeneses succeed.
We chose to experiment on sociogenesis for three reasons:
1. The first is related to the phenomenon of sociogenesis per se. The first occurrence of
emergence at the social level is to be found in the generation of the society itself. It
would be unnatural to study transformations of the social organization through the
impact of population increase, environmental variations or experimental manipulations
without first paying attention to the early stages of the society, which obviously
condition all its subsequent evolution.
2. The second reason relates to the validation of the model. Validation requires the
comparison of data obtained from the observation of natural societies with data from
the corresponding simulated ones. Sociogenesis provides us with at least two ways to
estimate the validity of the model, as the comparison can be done with two different
sets of data: demographic data, i.e. the number of each kind of brood and nestmates at
each stage of the development of the societies; and behavioural data, i.e. the individual
behavioural profiles of the ants and the functional groups they constitute during the
societies development. The success of natural sociogenesis under laboratory
conditions can be quantified easily by looking at the evolution of the colonys
demography. The analysis of behavioural data is more complex and, above all, timeconsuming. Moreover, detailed results about the sociogenesis of natural ant colonies
and their demographic evolution are already available (Corbara 1991; Corbara et al.
1991). It is therefore possible to evaluate the success of artificial sociogenesis by
studying only the
Artificial societies
166
MANTA
167
on growing, which means that reaching around 20 workers constitutes the end of the most
difficult period in the sociogenesis process. Table 10.1 reports the results in terms of
success and failure. Failure cases are clustered in seven categories, which correspond to
the composition of the population when the queen dies.
In these experiments, the proportion of failures (78 per cent), appears to be close to
that observed for natural colonies bred under laboratory conditions (where 86 per cent did
not reach the 10 workers stage). The situations in which the foundation of the colony is
likely to fail can be identified by the fact that larvae are part of the population (larvae
were present in 90
Artificial societies
168
(a) many workers are available, because other tasks still have a good probability of being
executed. But in the early stages of the foundation when the queen is alone, she is
prevented from doing anything else, even keeping enough food to feed herself;
Composition
Eggs
Number
Percentage
8
6.06
Eggs, larvae
16
12.12
Larvae
27
20.45
24
18.18
Larvae, cocoons
16
12.12
Larvae, workers
10
07.58
1.52
93
90.29
103
78.03
29
21.97
132
100.00
Eggs, cocoons
Failures with larvae
Total number of failures
Total number of successes
Total number of experiments
(b) food is not often far from the brood, which congregates near humid places so that it
can be neglected for a long time; and
(c) food has not begun to run short, and so the queen is not obliged to stay outside the
nest (or near its entry) until more food is supplied.
We have also conducted successful experiments. And, given the constraints stated above,
we will see which emergent strategy is employed by the queen and the first workers to
succeed in the foundation of the colony. By emergent strategy we mean a global longterm behaviour that looks like a strategy from the point of view of the observer, but which
has not been coded into the behaviour of the agents.
MANTA
169
Artificial societies
170
1. The dynamics of the sociogenesis process in natural colonies can be simulated by the
interactions of artificial agents provided with a small behavioural repertoire and a
simple mechanism of behaviour reinforcement. It does not mean that ants really
behave in this way; it just means that a stimulus/task model of behaviour at the
individual level, even if it does not account for the plasticity of real ants, is able to
gen-
MANTA
171
erate a complex pattern at the colony level which does account fairly well for the
development of ant colonies.
2. The whole demographic evolution of the society is generated without any centralized
control. The kind of birth regulation that emerges here is entirely due to local
behaviour and interaction between ants, or between ants and the brood, and is
conditioned by the constraints of the environment, as we shall see below.
Artificial societies
172
MANTA
173
Artificial societies
174
MANTA
175
Artificial societies
176
three workers. What a hierarchy provides is the possibility of switching dynamically from
the former to the latter. However, the lack of success of polygynous sociogenesis should
not be surprising. Our artificial queens have been provided with a behavioural repertoire
compiled from the observation of the queens behaviour in a monogynous species. Even
if the real queens can exhibit enough plasticity to react in a proper manner to a
polygynous situation (and that is what we are going to test by performing natural
experimentssee Note 4), the behaviour has not yet been studied. Hence, our simulated
queens are not likely to adopt behaviour adapted to this somewhat artificial situation.
Polygynous sociogeneses with a repulsive behaviour
We can now draw together hypotheses on the behaviours that are necessary for making
polygynous sociogeneses as efficient as those observed in other species. The first
hypothesis comes from The raulaz et al. (1990), who explain that, in the case of bumblebees, there is a link between the rank of an individual in the society and its mobility
inside the nest. According to them, a dominated bee will be less mobile than the bees that
dominate it. The second hypothesis comes from observation of the spatial positioning of
ants within the nest. When laying eggs, queens try to put them as close as possible to the
most humid places in the nest. Because of the position of the humidity, the eggs
aggregate in the top-left room. Eggs placed outside this area die more quickly than the
others. The idea we had was to combine these two hypotheses by providing the queens
with new behaviour, triggered directly by a stimulus they propagate, which makes them
flee the gradient associated with that stimulus. The consequence is that queens will now
repulse each other. Our assumption was that a queen already engaged in laying eggs
would push the other queens out of the brood room, thus resulting in a differential
positioning of the brood inside the nest. Being rejected from this room, the queens would
then lay eggs in other places, with less chance for them to grow old4. This would result in
turn in differential evolution of the lineages and reduce the demands of the brood,
converting these queens into worker-like ants.
Despite the plausibility of this chain of reasoning, our expectations were
Composition
Percentage
5
15.15
28
84.85
1 queen
20
71.43
2 queens
10.71
3 queens
10.71
Number
MANTA
4 queens
177
7.14
33
100.00
not fulfilled or, more precisely, not fulfilled in the way we had imagined. Quantitatively,
the experiments with these new queens were more successful than the previous ones with
either monogynous and polygynous sociogeneses (see Table 10.2). The proportion of
failures dropped to 15 per cent compared to 76 or 94 per cent.
However, the qualitative results were not what was expected. More than 71 per cent of
the successful experiments concluded with only one queen, the other three dying. And if
we look more closely at these societies, we discover that the deaths of the three queens
occurred in the first stages of the society (see Figure 10.11).
We observe that:
In the societies that end up with more than one queen, there is little spatial
differentiation among the queens. A queen occupying the brood room for laying eggs
or taking care of them does not always stay there. This is essentially because, being
closer, she is more exposed to the larval demand for food than are the other queens
and this obliges her to leave the room. While she is outside, the other queens can lay
eggs or carry their eggs into the brood room. Nevertheless, the repulsive behaviour
does seem to have a powerful impact (although not the one we were expecting) on the
success of the sociogenesis. It seems to allow a queen to complete a task without being
disturbed continuously by the other queens.
Explaining the success of societies that end up with only one queen is more difficult.
As a matter of fact, most of the simulation runs look similar to the monogynous
sociogeneses. We have, however, observed that:
(a) the most difficult stage of the foundation process is when the first
Artificial societies
178
Conclusion
All the results presented here are preliminary (except those on monogynous
sociogeneses) and we are still conducting experiments with alternative hypotheses in
order to understand the mechanisms underlying the generation of social structure within
ant societies.
The results that we are waiting for are those on the division of labour within the
colonies. These are still being analyzed and examined, but because they represent a huge
amount of work, they will not be available for some time. We need them to begin
experiments on sociotomies (Lachaud and Fresneau 1987)the splitting of a colony into
two or more subcolonies made of different functional groupsin order to observe how
these subcolonies produce the specialists that are lacking. We also want to conduct
studies on the relationships between social structure and spatial organization in the nest
(see Fresneau et al. 1989).
We conclude with a word on the simulation we have conducted on the generation of
hierarchical structures. Apart from the interest of such a simulation in itself, the
underlying problem it emphazises is the apparently unnatural relationship between
competition and co-operation. As it is programmed, the behaviour of the queens makes
them compete with each other without any kind of collaboration. However, the results
obtained with the polygynous sociogeneses (and especially those concerning the colonies
MANTA
179
that terminate with one queen) can be interpreted easily in terms of co-operation between
the queens. In most cases it is even possible to describe the evolution of these colonies by
using notions such as altruism or social sacrifice that are currently employed by
sociobiologists to describe similar phenomena among social insects. What we want to
underline is that the notion of co-operation is a very fuzzy one and that many processes or
behaviours viewed as being co-operative may be obtained by the competitive interplay
between individuals and the constraints generated by their environment.
Acknowledgements
We are very grateful to Michel Denis, IPT-Universit Paris XIII, for having allowed us to
use 15 IBM computers which were the hosts of our simulated ant colonies for several
weeks. We also wish to acknowledge the contribution of Steffen Lalande in programming
the Pascal version of MANTA during the first year of his doctorate; Nicolas Czternasty
for having patiently conducted most of the experiments on hierarchy; and Philippe Louise
for having discovered the importance of birth control, even if it was unplanned. We thank
them for their wise advice and stimulating discussions.
Notes
1. This process is detailed in (Drogoul et al. 1992a). It consists simply in incrementing the
weight of a task when it has been activated.
2. In fact, the inverse of this motivation: when the threshold is low, the agent is considered to be
very motivated.
3. A hundred experiments have begun on polygynous sociogenesis within a monogynous
species at the Ethology Laboratory of Paris XIII.
4. It has to be noted that we do not pay attention to the phenomenon of inhibition that has been
reported to occur in many species of ant. This phenomenon makes a queen inhibit the
reproductive capacities of the other potential reproductive females by propagating a special
pheromone. We shall probably consider it in our future experiments on hierarchical
organization. The problem is that we would have to modify the model of behaviour of the
agents in order to implement the inhibition of behaviours along with their activation.
5. A queen with a high weight would have been repulsed more frequently than one with a low
weight and could then be considered to be dominated.
Chapter 11
Emergent behaviour in societies of
heterogeneous, interacting agents: alliances
and norms
Nicholas V.Findler and R.M.Malyankar
This chapter deals with alliance formation and norms of behaviour in distributed
intelligent systems. An alliance is a social entity composed of intelligent agents that have
made a deliberate decision to join forces to achieve alliance goals. Norms are defined as
limitations on agents, such as legal constraints on individual actions. These phenomena
are considered to be emergent behaviour in the context of distributed intelligent systems.
Emergence is defined as the appearance of patterns of behaviour that are discernible only
at higher levels of observation. Multiple levels of emergence are possible. The aim of the
research is to arrive at descriptive and prescriptive theories of group behaviour.
The chapter begins by describing the basis for alliance formation in terms of the goals
of individual agents. Procedures for computing the similarity of agent goals and
procedures for alliance formation are informally and formally described. This is followed
by a description of a mechanism for implementing norms and for the evolution of norms
over time. Finally, experiments in progress using these procedures are described in the
context of international power politics.
It is hoped that these studies will contribute to the understanding of human
organizations, including international political action, national and international markets,
industrial problems, co-ordinated efforts, and other social situations, both natural and
artificial.
Introduction
Distributed Artificial Intelligence (DAI) is, in general, concerned with how the members
of a group of intelligent agents perceive and reason about the environment, perform,
interact, collaborate and/or compete with each other (Bond & Gasser 1988b, Durfee et al.
1989, Findler & Lo 1986, Findler 1987, 1990, 1992). This chapter describes efforts
towards developing a system to study collective phenomena in the emergent behaviour of
heterogeneous agents. It uses the concepts and methods of computer science and artificial
intelligence as well as those of organization theory, sociology, social psychology,
political science, economics and management science.
This research deals with communities, rather than collections, of intelligent agents.
The agents may have different capabilities and resources, knowledge bases, goal
structures, value scales, representations and viewpoints of the environment, social and
organizational attitudes, methods of reasoning and decision-making, commitments and
belief structures. The aim is to obtain descriptive, explanatory and prescriptive theories of
group behaviour. These theories would tell us how individual members form and dissolve
181
non-permanent alliances that share some of the members goals, act under joint
constraints (principles), exchange messages, make decisions, plan actions, collaborate
and compete with other alliances. The ultimate aim is to learn why the total behaviour is
more than the average of the individual behaviours. The term alliance is used to refer to
a stronger form of co-operation than generally found in the DAI literature (Davis &
Smith 1983, Durfee et al. 1987, Findler & Lo 1986, 1991, 1993, Findler & Stapp 1992,
Findler & Sengupta 1992).
The alliances are social entities subject to certain conventions and guidelines, called
norms. The norms define the legal limits of goals and actions, and can be changed if
enough members of the alliance so decide. However, there are certain constraints and
limitations, called meta-norms, that determine in what direction and how far the norms
may change.
Alliances and norms are regarded as forms of emergence in distributed intelligent
systems. Emergence is defined as the appearance of patterns at higher levels that are not
apparent at lower levels. Multiple levels of emergence are possible; patterns of behaviour
at intermediate levels may in turn constitute emergence at higher levels. The concept of
emergence adopted here is somewhat broader than the usual definition, which requires
that agents be unaware of the existence of potential for emergence (see also Chapter 8).
Agents in distributed intelligent systems may, if they possess sufficient intelligence, be
aware of the potentiality of emergence and strive to enhance (or destroy) emergence. One
example of such behaviour is a stock market, where the individual agents (brokers) are
aware of the potential for cumulative consequences of many similar individual acts; for
exam-ple, a rush to sell or buy stock in a particular company has effects on its price, and
sometimes on the market as a whole, that are not accounted for in terms of the effects of
individuals actions. In such cases, the emergence is within the reasoning horizon of
individuals, since they are aware of the possibility of emergence and may try to influence
it. However, it is outside their control horizon, since (in most cases) agents cannot
influence it by their individual actions alone. Such phenomena, which are not controllable
by individuals, are also considered in this chapter to be instances of emergence. Alliances
and norms are emergent phenomena given this definition, as they are (in general) not
under the control of any one agent. Also, alliances exhibit (in some respects) united
action, where the individuals give up some or much of their independence in return for
safety in numbers.
The concept of the alliance replaces that of a single agent as the unit actor with a
single locus of control and intention. One must keep in mind the following features that
distinguish alliances from other forms of co-operation:
1. Alliances may differ from one another in all possible characteristics and have no a
priori motivation to collaborate.
2. If two or more alliances decide to collaborate, it is based on their momentarily shared
goals, and on the accuracy and timing of their network perception (their perceptions
of other alliances).
3. The goals of an alliance may be self-generated or induced by the environment.
4. The members of a single alliance or different alliances may assume a temporary,
hierarchical or flat organizational structure if it serves their purpose and their internal
power conditions.
Artificial societies
182
5. The membership and the norms of an alliance may change dynamically in response to
the changing environment and the individual agents mentality.
6. The actions of an alliance may affect the others adversely if their goals are partially or
totally exclusive.
7. The costs and benefits of an alliance depend on the actions of the other alliances as
well. It may, however, be a non-zero-sum environment.
8. Conflicts arising between alliances may (but need not) be resolved through negotiation.
9. The malfeasance of an alliance (deliberate or accidental transgression of the norms) is
to be discovered and reported by others.
10. When the goal and/or action of an alliance are found to violate the current norms,
their expected value to the social entities in question is calculated. If this is greater
than a certain threshold value, a suggested change to the norms (allowing the goal or
action in question) is checked vis--vis the meta-norms, and, if found to be feasible,
put to a vote. A negative vote by the majority (or plurality) can be followed by
negotiations leading to an acceptable consensus.
The features listed above are observed in many alliances in international relations,
legislation, economic behaviour, and other areas of social activity. Not all the features are
necessarily present in all alliances.
Our studies are intended to enhance the understanding of human organizations, the
social and organizational role of knowledge and meta-knowledge, national and
international markets, industrial problems, the effect of conflict and competition, tradeoffs in co-ordinated efforts, and computer tools for the control of co-operation and
collaboration.
The rest of this chapter concentrates on the concepts of alliance and norms. The next
few sections present formal and informal descriptions of procedures for alliance
formation, dissolution and concerted action by the members of an alliance. These sections
are followed by a formal definition of norms and procedures for generating, evolving and
applying norms. Finally, the use of these procedures in a single social domain is
described.
Goal structure
The basis and motivating force of action is the description of the goal structure of agents
and alliances. The following formal description of goal structures requires the definition
of a similarity operator, a binary operator that accepts as its operands two goal structures
and returns a numerical value characterizing the degree of similarity between them. In
addition, the formal definition of a goal structure must be powerful enough to allow for
the description of complex goal structures, that is, complex Boolean expressions using
the AND and OR connectors between clauses. Accordingly, the structure of a goal is that
of an AND/OR tree. The formal definition of the goal structure follows:
183
Artificial societies
184
necessarily domain-dependent, since the degree of matching between two goals depends
on the semantics associated with the descriptions as well as the descriptions themselves.
For example, G3, stated as the red ball is beside the blue box should match G1 to a
lower degree than G2, as the relationship between the two objects is different. (Here we
assume that position is more important than colour. If the opposite holds, the result of the
comparison changes.)
The steps in computing the similarity between two goals G1 and G2 are:
1. The basic_match function is applied to the descriptions of G1 and G2. If this results in
a non-zero value for the degree of match, the value is returned as the result. If the
match value is 0 and both G1 and G2 are atomic goals (that is, have no subgoals), 0 is
returned.
2. If Step 1 fails to return a value, the goals G1 and G2 are compared with subgoals of
each other. This handles the situation where one goal is a subgoal of the other. For
example, have a 600-ship navy may contain expand Portsmouth Naval Shipyard
as one of its subgoals. Each of these may be a top-level goal for different
Representatives in the US Congress. If this step results in a non-zero match value, the
value obtained is returned.
3. If Step 2 fails to return a value, the subgoals of G1 are matched with the subgoals of
G2, calling the similarity computation recursively, that is, starting over with Step 1 for
each pair of subgoals. The values obtained are combined to give the match value for
the parent goals. The form of combination is based on whether the parent goals are
AND or OR goals, that is, whether the subgoals are conjuncts or disjuncts.
4. If all steps above fail, 0 is returned.
A more formal description of the similarity computation is beyond the scope of this
chapter. Its definition tries to mimic the intuitive estimates of similarity between goals in
most social situations, which are based on a deep understanding of the implications of
various goals and their require-ments in a social environment.
The definition of this operator requires precise descriptions of individuals goals.
These descriptions must be written to include explicitly the dependence of the goals on
agents; for example, if both agent As and agent Bs goals are to maximize their
individual utilities, the goals must be expressed in forms that translate as maximize As
utility and maximize Bs utility, respectively, and not simply as maximize utility.
The first formulation allows the potential conflict between the two goals to be
recognized, while the second does not. The alternative is to make the basic_match
function smart enough to recognize such ambiguities and to make it consider the
identities of the owners of the goal structures when computing the similarity of goal
descriptions. These restrictions mean that a situation where one agents utility is other
agents disutility is taken care of.
The rationale behind this approach to defining similarity is as follows. The
computation of similarity depends on two things: first, the structural similarity between
two different goal structures; and secondly, the degree of semantic match between the
descriptions of those goals (in an artificial language invented for the domain, if
necessary). The above formulation of the operator separates the semantic content of
similarity, which is necessarily domain-dependent, from the structural content, which is
not. Accordingly, application of this operator to another domain requires the replacement
185
of the semantic content (that is, the redefinition of the basic_match function), while the
structural content can be carried over without change. Of course, this carrying-over
requires that the goals in both domains can be expressed as goal trees. This requirement is
not overly restrictive as the tree representation of goals is general enough to cover most
goal structures. This approach to defining the similarity between goals is, we believe, as
generic as possible for the problem.
Importance assignment
The importance of a goal is a number between 0 and 1 (inclusive) that gives the relative
importance of the goal to the agent or alliance to which the goal structure belongs.
Importance values are assigned to top-level goals based on the agents or alliances
internal considerations as well as the external environment. The distribution of top-level
importance to subgoals is described below. Each case corresponds to one way of dividing
a root or intermediate goal, as described above.
GX
importance(X)=importance(G)
GA1 A2
importance(A1)=importance(A2)=importance(G)
GA1 A2
importance(A1)/importance(A2)=size(A1)/size(A2)
max(importance(A1), importance(A2))=importance(G)
The first case simply expresses the importance of an atomic goal in terms of the goal
from which it is derived by means of the production rules that define a goal structure. The
second case means that all AND subgoals of a goal are as important as the goal itself,
since all are necessary to achieve it. The third case is for OR subgoals, and does two
things: first, it prioritizes the subgoals in inverse order to their size and hence their
complexity; and, secondly, it specifies that the importance of the highest-importance
subgoal shall be equal to the importance of the parent goal. These conditions taken
together mean that achieving the simplest subgoal has the same importance as achieving
the goal G, while achieving more complex goals has a lower importance, since they are
alternatives to the simplest subgoal.
The importance is used in the basic_match function to allow fine tuning of the
similarity based on the degree of importance each agent attaches to the goal or subgoal. It
is also used in selecting alliance goals.
Artificial societies
186
structure out of the goal structures of individual members. This chapter defines this
procedure in terms of two subprocedures: one to select the most mutually beneficial
member goal to be the initial alliance goal, and another that constructs an alliance goal
structure around the selected initial goal. These procedures are now described
informally.
The procedure to select the initial alliance goal is:
1. Choose arbitrarily any two agents from the alliance and combine all their goals (toplevel goals as well as subgoals) into a rank-ordered set. The rank ordering is based on
the degree of similarity between goals. Goals that are similar to many other goals are
ranked higher than goals that are unique. The reasoning behind this method of
ordering is that alliance goals are most likely to be those goals that contribute to
furthering the aims of most members of the alliance.
2. Once a rank-ordered set has been created, the goals of other members of the alliance
are added to the set one by one, using the same ranking procedure. In other words, for
each goal or subgoal of an alliance member, a rank is computed based on how many
goals it matches, either partially or completely. Furthermore, goals already in the set
have their rankings updated, based on the degree of match between them and the new
goals being introduced.
3. The result of Step 2 is an ordered set of goals that is independent of the order in which
member goal structures were processed. This set contains all the goals in all the goal
structures of alliance members. The set is ranked by mutual benefit, that is, the first
goal in the set is the one that will benefit the members the most, by virtue of matching
many goals in the goal structures of many members. This goal is selected as the basis
for constructing the alliance goal structure.
All the agents determined to join a particular alliance do so simultaneously, after which
the new goal structure of the alliance is computed. Working in this way, the membership
of the alliance does not depend on the order in which individual agents join it.
Whenever an alliance is formed, or when it is reorganized, an alliance goal structure
must be constructed from the goal structures of individuals or coalescing alliances. This
construction, which is carried out by the second procedure, starts with the most mutually
beneficial goal and adds subgoals and parent goals of this goal to the alliance goal
structure one by one. While a goal is being added to the alliance goal structure, the
importance of the goal to the alliance is calculated. The procedure that implements these
ideas is as follows:
1. Let the highest-importance alliance goal, selected by the first procedure, be denoted by
gx. Add this goal to the (initially empty) alliance goal structure. It is assigned an
importance value of 1 (the highest possible).
2. For all members of the alliance:
(a) find the location, if any, of the selected alliance goal gx in the agents individual
goal structure; and
(b) add each subgoal gy of gx in the agents goal structure to the alli- ance goal
structure, as a subgoal of gx.
187
3. Repeat Step 2 for each non-atomic goal in the alliance goal structure, until there are no
non-atomic subgoals left. In other words, all descendants of the original goal are added
to the alliance goal structure, until the subtree rooted at gx is fully developed.
4. Add the parent goals of the originally selected goal gx in a similar fashion, until either
there are no more goals to be added or a specified importance cut-off has been
reached.
5. Assign importance values to the constituents of the alliance goal structure, starting
with the original goal (already assigned an importance value of 1):
(a) descendant goals of the original goal are assigned importance values calculated
from the type (AND/OR) of their parent goal, as discussed in the previous section;
(b) ancestor goals of the original goal are assigned importance values based on their
ranks in the ordered set computed by the first procedure described; and
(c) any goals in the alliance goal structure that have not already been assigned an
importance value are assigned values based on the importance values attached to
their parents, as described in the previous section.
This procedure constructs a goal structure that is the composite of the goal structures of
the alliance members and is based on the most mutually beneficial goal, in the sense that
no other goal has a greater importance value than this goal.
It is possible that two identical higher-level goals have somewhat different goal
structures. For example, within the context of legislative action, representatives from
different states may have the same top-level goal (such as build a 600-ship navy) but
different subgoals (expanding naval shipyards in their respective home states). In this
case, both subgoals will be placed in the alliance goal structure, and the potential
contradiction between the two subgoals will be resolved within the alliance by an
accepted voting process, so that the alliance goal structure may remain conflict free.
(Recall that the members deciding to stay in the alliance are not allowed to pursue their
own goals.)
Artificial societies
188
endfor
TotalUsefulness TotalUsefulness+Usefulness(gx)
MaxUsefulness MAX(MaxUsefulness, Usefulness(gx))
endfor
4. if (TotalUsefulness > TotalThreshold)
return TRUE
else
return FALSE
endif
Steps 1 and 2 initialize the goals and the potential helpfulness (MaxUsefulness) of the
alliance to the agent. Step 3 compares each goal in the agents goal set to each goal in the
alliances goal set and calculates two values: the overall potential helpfulness
(TotalUsefulness), defined as the sum of helpfulness for all goals; and the maximum
potential helpfulness (MaxUsefulness), the maximum extent to which any one of the
agents goals is helped by the alliance. If either of these is above a threshold, the agent
decides to join the alliance. This amounts to using two criteria: (i) the overall expected
utility of joining the alliance (TotalUsejulness); and (ii) the expected utility for a single,
very important goal (MaxUsefulness). The extent to which one goal helps along another
is obtained using the facilitates calculation, described in the next section.
Calculation of facilitation
This calculation is used to determine the utility of one goal in facilitating another goal. A
goal is said to facilitate another goal when achieving the first goal helps the achievement
of the second. For example, achieving a subgoal facilitates the achievement of that
subgoals parent. The degree to which one goal facilitates another is measured in terms of
the number of steps required to get from the first goal state to the second. The algorithm
is constructed so that the facilitation is a value between 0 and 1, with 0 corresponding to
no help at all (or insignificant help) and 1 corresponding to direct logical implication:
facilitates(g1,g2) degree to which goal g1 facilitates goal g2
This is calculated by computing successive action-closures of g1 (the sets of propositions
that become TRUE after successive possible actions), starting with the state in which g1
holds. The computation of action closures is ended after a limited number of steps, at
which point the facilitation value can be assumed to be low enough to be ignored. (In
principle, facilitation could also have negative values, corresponding to a counter-
189
effective goal pair. This ramification is not explored in this chapter.) The algorithm for
calculating F, the value of facilitation, is as follows:
1.
F 1.0
2.
then
return F
else
F F/2; let CL {g1};
endif
3.
4.
return 0
Some explanations are needed. The set CL is the action-closure of g1; that is, the set of
states which can be reached from g1 with 1, 2, 3,... successive actions. Steps 1 and 2
initialize the values and check whether g1 in the current state directly implies g2 (that is,
without any actions needing to be taken by agents). Informally, step 3 tries to find out
how much effort is needed to achieve goal g2 starting from g1 (and the current state of the
environment). This effort is measured in terms of the number of steps required to achieve
g2. The rationale is that the more steps are required to achieve g2 (given g1 in the current
environment), the less g1 facilitates g2. The division factor 2 by which facilitation values
are reduced in successive steps and the cut-off threshold, LowThreshold, below which
facilitation is considered to be insignificant, are both arbitrary values.
For each pair of agents, calculate the similarity between their goals.
This gives a proximity matrix for all known agents.
2.
Artificial societies
190
191
by noticing that they do in fact have a common goal: defeating the other agents. The
goal-based alliance formation described in this chapter does not require that the agents
goals stay the same throughout; the individual goals are allowed to change in response to
changes in the environment, actions of other agents, improved knowledge, or negotiation
with other agents.
Artificial societies
192
made, the norms are clearly no more than a set of rules that are applied to actions and
goals, and the norm-related computation is a kind of expert system selection mechanism
(filter) applied to the actions/goals (Giarratano & Riley 1989).
For the purposes of this chapter, norms are defined only within the context of a
particular alliance, that is, a particular norm applies only to the individual members of a
particular alliance. Meta-norms are constraints and limitations that are functionally
similar to norms, but apply over the entire environment; that is, they apply uniformly to
all the agents and alliances in the system. Meta-norms correspond, for example, to
humanitarian, ethical and moral standards in the real world. Norms correspond to
regulations aimed at improving efficiency and adherence to the meta-norms.
One significant difference between the application of norms/meta-norms and an expert
system is that the satisfaction of a norm or meta-norm by a set of goals need not mean
that it is in fact followedit might be violated, or the agent might choose to release itself
from the norm by leaving the alliance. The violation of a meta-norm may come about
through an explicit decision by the alliance to violate it when the payoff is high enough,
where the alliance is willing to risk discovery and also risk the penalty for the violation.
Norms, on the other hand (which apply to members of an alliance) are not violated by the
members. Instead, agents that find the costs of staying in the alliance to be too high may
choose to depart from the alliance, paying the costs of departure, if any. Another
difference from expert systems is that the norms apply only in certain circumstances, that
is, (in general) only to members of a single alliance.
The rest of the discussion assumes only one kind of norms and meta-norms: those that
forbid certain actions/goals.
Norm generation
The purpose of norms is to enable more efficient functioning of the distributed system
and to prevent harmful events. Efficiency and harm in this context are measured by
criteria that are important to the particular system. For example, in a legislative context,
one of the criteria for determining efficiency is the time taken to pass controversial
193
legislation. Harmful events are also defined in a context-sensitive manner. In the real
world, harm is measured in terms of transgressions of humanitarian, ethical and moral
standards. For example, the destruction of an agent (say, by reducing all its resources to
zero) may be considered to be a harmful event in some distributed intelligent systems.
This is implemented in practice by providing a list of harmful events (in much the same
format as the description of a norm) while setting up the system. In other words, the
concepts of harm and efficiency are domain-dependent. They may be measured in terms
of their effects on the payoffs to individual agents. Because in the simple case considered
in this chapter, the agents are selfish individuals that try to maximize individual payoffs,
the sum of individual payoffs constitutes the global payoff.
Thus, norms must arise from experience; that is, they must be formed after a harmful
event or inefficiency is detected. In other words, a norm is in a sense a new concept that
must be learned from examples or non-examples/ counter-examples. This means that a
record of actions/goals and their consequences and payoffs must be kept. Whenever a
harmful event occurs, the log is examined and a norm constructed (for example, as a
conjunction of the actions/goals immediately preceding the harmful event). Deciding
which of the immediate precedents of a harmful event caused the event is a difficult
problem in itself, and will be simplified for the purposes of this research by depending on
an omniscient black box to construct a norm from the log by deriving causal relations
between events and their precedents. This is adopted as a conceptual simplification; any
implementation of this black box is necessarily domain-dependent. Note that in this
context, omniscience does not mean infallibility. The black box may (and does) come up
with norms that are wholly or partially wrong. Norms are suggested by an individual
agent (this means that each agent is capable of deriving and evolving norms). A norm,
once suggested, may be negotiated with other agents and subjected to a voting procedure
to decide its acceptance. The only method of norm generation used at present is direct
forbidding of the action in question. This method is more general than it seems, and is
often seen in human affairs (for example, laws are often devised that directly forbid
specific actions that are regarded as harmful to society).
The norms generated in this phase are evolved, ideally towards perfection, by the
agents. Figure 11.1 shows the process whereby norms are generated.
Artificial societies
194
B0
2.
3.
195
return B
Discovering violations
Violations of norms are discovered by an agent during its perceive phase, by matching
the most recent actions of other agents from the same alliance or actions of other
alliances, to the existing norms (in the case of members of the same alliance) and metanorms (in the case of other alliances). Penal- ties are levied on violators in the form of
(a) downgrading the importance of its goals in the alliance goal structure, perhaps after
holding a referendum;
(b) expulsion from the alliance; and
(c) retaliation directed at the offending agent or alliance.
A domain of application
The first chosen domain for experimentation is international politics, in particular, postSecond World War international relations in North Africa. Within this domain, we study
the phenomena of alliance formation and dissolution related to the political alliances
between nations of the region and the superpowers interested in the region. The general
goals of these alliances can be identified as follows:
1. Deterrence of aggression by other countries/alliances.
2. Defence against military moves by other countries/alliances.
3. Alliance unity in political and military action.
4. Enlargement of the alliance.
5. Reduction in the membership of competing or hostile alliances.
6. Non-proliferation of nuclear and certain other weaponry.
7. Control of escalation in crises.
8. Stabilization of crises.
9. Stabilization of arms race in general.
10. Control of costs.
11. Elimination of the possibility of surprise attacks.
12. Reduction of extremism and terrorism.
13. Gaining economic advantage in trade.
14. Influencing the public opinion in the other alliances.
15. Adjustment of the internal political pulls within each nation to increase the influence
of the group currently in power.
Of the above, some goals (deterrence, defence, etc.) are noted by Kugler (1990). Within
these general goals, each alliance or agent (nation) will have its own goals, which may
conflict partially with the goals of another agent or alliance.
Artificial societies
196
2. Would these alliances form and evolve just the same as they have in reality?
Since the scope of the domain is far too large for reasonable modelling efforts, our
first experiment is restricted to a small number of major events rather than modelling the
entire domain over the period of interest.
An experiment
The experiment described in this chapter concerns the SovietLibya rela- tionship. For
the purpose of this experiment, the salient points are:
1. The goal structure of both parties.
2. The value of each party to the other.
3. Considerations determining the selection of the other party instead of other candidates.
For both parties, the goal structure consists of a main goal and auxiliary component or
subgoals. For each party, the overall goal is to enhance its own status. The sub- and
auxiliary goals derive from this overall goal. For the Soviet Union, the auxiliary goals
deriving from its root goal (enhancing its own power) are: (a) enhance status without
regard to ideology; and (b) enhance status with regard to ideology. (History has shown
that these two were at times quite distinct and even in conflict.)
The first derives from superpower compulsions and the second from the ostensible
ideological background of the Soviet Union. These second-level goals can be further
subdivided into gaining (a) military, (b) political, and (c) economic advantage over the
Western powers.
Of these, only the political advantage subgoal is considered to apply to the secondlevel goal of gaining advantage with regard to ideology. (This restriction is introduced for
purposes of simplicity, and is not considered a general constraint.) These third-level goals
are composed of the following subgoals:
1. Gain military advantage. This can be further subdivided into: (a) gaining military
facilities; (b) improvement of military discipline among allies; and (c) getting a
positive military effect on the neutral countries.
2. Political advantage. This can be further subdivided into: (a) gaining political friends,
which includes favourable votes in international fora; (b) improving political
coherence in the Soviet camp; and (c) projecting a positive political image to neutral
countries.
3. Economic advantage. This can be further subdivided into: (a) gaining economic
benefits, such as raw materials at favourable prices; (b) improving economic
coherence in the Soviet camp, with a view to increasing trade under favourable
197
conditions; and (c) projecting a positive economic influence on neutral countries, with
a view to drawing them into its sphere of political influence.
For Libya, the goal structures are similar, although the relative importance given to each
component varies and the goals specific to superpower concerns are removed. The
descriptions are also changed to match Libyas concerns, for example, expanding
economic influence in the Soviet camp becomes expanding economic influence in the
Arab bloc and in African countries.
These goal structures are encoded as in Figure 11.2. Here, each line represents a goal
or subgoal. The structure of the goal tree should be clear from the indentation. The first
word after the left parenthesis indicates whether the goal is an AND goal (a conjunction
of its subgoals), an OR goal
(or enhance status of s-u)
(and regardless of ideology)
(or military advantage)
(atom military facilities vs western powers)
(atom improve military discipline soviet-camp)
(atom positive military effect on third world)
(or political advantage)
(atom gain politioal friends)
(atom improve political discipline soviet-camp)
(atom positive political effect on third world)
(or economic advantage)
(atom gain economically profit)
(atom improve economic discipline soviet-camp)
(atom positive economic effect on third world)
(and with ideological coverage)
(or political advantage)
(atom support favorable organizations)
(atom advance propaganda aims)
(atom cause factions in opposing parties)
Artificial societies
198
Related work
Some research on coalition forming has been described in the literature on group
behaviour (for example, Mokken & Stokman 1985). In relation to coalition forming in
the legislative and international domain, on the other hand, there is significantly more
literature, especially concerning voting coalitions in legislative assemblies (for example
Cigler & Loomis 1986, Davidson & Oleszek 1985, Mokken & Stokman 1985, Oleszek
1989). Most of the literature appears to be aimed at describing the phenomenon of
coalition formation. Our research tends to be prescriptive rather than descriptive in that
we attempt to arrive at a framework for the process of alliances and coalition formation in
algorithmic terms rather than capture the individual features of coalitions/alliances in a
particular domain.
199
Acknowledgements
The authors wish to thank the anonymous referees for their comments and suggestions.
Part III
The foundations of social
simulation
Chapter 12
Kin-directed altruism and attachment
behaviour in an evolving population of
neural networks
Domenico Parisi, Federico Cecconi and Antonio Cerini
In early human societies social groups were mainly based on space. A social group was a
set of individuals who tended to live close together in the same local environment.
Physical nearness made it possible to interact socially and to develop social structure
inside the group. It is only with the development of transportation technologies and,
especially, the huge development of communication technologies in the twentieth
century, that space has lost some of its importance in determining who is interacting with
whom and, therefore, the nature of human groups.
A second property of early social groups was their biological basis. The social
sciences have a tendency to ignore the biological infrastructure of human societies but we
know that these societies have evolved from a state in which human groups were
essentially families, i.e. sets of individuals who were genetically related or related
because of sexual reproduction (Johnson & Earle 1987).
In his 1992 presidential address to the American Sociological Association (Coleman
1993), James S.Coleman remarks that the discipline of sociology was born at a time when
Western societies were changing rapidly from what he calls primordial societies, based
on locality and kin relations, to modern societies, in which social relationships tend to be
independent of space and biology. Sociology has concentrated on modern societies and it
has embraced the point of view where in order to understand X it is not necessary to
reconstruct how X has come to be X, i.e. its origin and development.
Our preferences go to the opposite (genetic) viewpoint (see Parisi in press), according
to which contemporary human societies cannot be understood unless one examines how
they developed from earlierin fact, even non-human primatesocieties in which space
and biology had a more important role. The simulation of human societies, for us, is the
simulation of an evolutionary or historical process of change. More specifically, we are
concerned with simulating how human groups changed from being small-sized groups
largely based on space and kin relations to being progressively larger groups in which
other causal factors beyond space and biology kept the group members together.
In this chapter we examine two elementary forms of social behaviour which appear to
play a role in creating simple social groups in organisms, both human and non-human,
i.e. attachment behaviour and kin-directed, genetically-based altruism. Attachment
behaviour is the behaviour of an organism that tends to keep it near another organism.
Hence, attachment behaviour is one basic mechanism for maintaining space-based social
groups. Kin-directed altruism is the behaviour of an organism which increases the fitness,
i.e. the reproductive chances, of a kin while decreasing the fitness of the organism
Artificial societies
202
exhibiting the altruistic behaviour. Attachment behaviour and altruism are related and
are, in fact, complementary behaviours. If it is the case that the organism which is the
potential recipient of an altruistic act can only obtain the advantages emanating from the
act if it is near the altruist, then there will be a tendency for attachment behaviour to
emerge through evolution. In the simulations we shall describe below, offspring exhibit a
tendency to follow their parents in their wanderings in search of food because offspring
can only receive food (and survive) if they remain close to their parents, who feed them.
It is more complicated to explain how altruistic behaviour may emerge and be
maintained in a population of organisms. Since altruistic behaviour by definition
decreases the fitness of the altruist, individuals with altruistic tendencies will tend to have
fewer descendants than do non-altruists and therefore they should in the long run
disappear from the population. Kin-selection theory (Hamilton 1964, Trivers 1985) has
been introduced to explain why, contrary to this expectation, genetically-based altruism is
observed in many organisms. According to this theory, altruistic behaviour will be
maintained in a population if it is directed towards ones own kin. The fitness of the
altruistic individual will be decreased by the altruist act but the fitness of the kin
individual who is the recipient of the act will be increased. Since kin individuals are
genetically similar and are therefore likely to exhibit the same behavioural tendencies,
altruistic behaviour present in the reproductively unsuccessful altruist will also be present
in the reproductively successful kin towards which the altruistic act was directed. Hence,
whenever the increase in fitness of the recipient of the altruistic act exceeds the altruists
loss of fitness, altruistic behaviour will emerge and be maintained in an evolving
population.
The present chapter is divided into several sections. In the next section we describe
some simple simulations whose purpose is to demonstrate that kin-directed altruism can
emerge through evolution and be maintained as a stable trait in a population of artificial
organisms. In the following section we describe more complex simulations of populations
of artificial organisms living in a shared environment. In these simulations not only does
genetic altruism of parents toward their offspring emerge, but also offspring develop a
tendency to follow, i.e. to remain in proximity to, their altruistic parents. Therefore, in
these simulations we observe the emergence of a simple kind of social group composed
of individuals that live in the same local environment and exchange simple acts of
following and giving food. In the final section, we discuss how to extend the range of
phenomena that can be simulated along the same lines.
203
must decide whether to act altruistically or egoistically with respect to its current social
partner. Each network has two input units, three hidden units, and one output unit. The
two input units encode the type of the current social partner which can be either a sister
(10) or a non-sister (01). A sister is another network which is an offspring of the same
parent of the network we are considering. (In all simulations reproduction is agamic, i.e. a
single parent generates one or more offspring.) The single output unit encodes whether
the network decides to give (1) or not to give (0) a piece of food in its possession to its
current social partner. At the end of life the networks with the largest number of pieces of
food (fitness) reproduce, while the other networks die without leaving offspring.
The initial population is composed of 100 networks divided into 20 groups of five
sisters (identical networks). Each group of sisters is assigned a different set of connection
weights but all the sisters in a group share the same weights. All networks live the same
number of cycles. At the beginning of life each individual is given the same number of
pieces of food. In each cycle the network has a different social partner randomly selected
from the population. Hence, at any given cycle the social partner may be either a sister or
a non-sister (with different probabilities). The network must decide whether to act
altruistically by giving its current partner one of its pieces of food or to act egoistically by
refusing to give food. If a piece of food is given, the donors food counter is decreased by
one unit while the social partners counter is increased by one unit. Otherwise, nothing
changes.
At the end of life, which lasts 100 cycles (i.e. 100 social encounters) for all
individuals, the 20 individuals with the highest number of pieces of food reproduce by
generating five copies of their matrix of connection weights, with random mutations
added to some of the weights. The other individuals die without leaving descendants. The
offspring networks constitute the second generation. The second generation also consists
of 20 groups of five sisters each but, unlike the first generation, sisters now have similar
but not identical connection weights because of the random mutations. The process
continues for 50 generations.
Since individuals reproduce as a function of the number of pieces of food they possess
at the end of life, one would expect individuals who always act egoistically, that is, who
never give away food, to emerge and colonize the entire population. These individuals
keep their own food and may occasionally receive additional food from altruisticallyacting individuals. However, as predicted by kin-selection theory, what is in fact
observed is a more complex type of behaviour (all results are based on five replications
of each simulation, each starting from a different seed of a random number generator).
Individuals refuse to give their food to non-sisters but they tend to give it to sisters. In
other words, they act egoistically towards non-kin but altruistically toward kin. The loss
of fitness that is caused by giving food to sisters is compensated by the increase in fitness
of the sisters who receive food. Hence altruistic traits, i.e. connection weights which
result in giving food, are maintained in the population via reproducing sisters (inclusive
fitness). However, altruism is exhibited only towards sisters. Non-sisters are treated
egoistically (Figure 12.1).
Artificial societies
204
205
changes its status and it becomes an adult. At the beginning of adulthood the individual is
endowed with a certain amount of energy which is equal to its parents current energy.
This energy is also decreased by one unit at each cycle. However, by capturing food in
the environment an adult can increase its energy. An adult generates a new offspring at
regular intervals (number of cycles) during its life. The individual is also asked regularly
whether it wants to give one unit of its energy to its offspring or it wants to keep all the
energy for itself. By deciding to keep its energy for itself (and not to give it to an already
existing offspring) an adult not only makes its own life longer but also increases its
chances of generating further offspring.
The behaviour in the environment of both adult and young individuals is controlled by
a feedforward neural network. Neither the networks architecture nor the networks
connection weights change when a young individual reaches adulthood. What does
change is what is encoded in the networks input (sensory) units. Each network has four
input units, six hidden units, and two output units. When an individual is young, two
input units encode the current position of the individuals parent relative to the young.
One unit encodes the angle of the parent (measured clockwise and mapped in the interval
between 0 and 1) with respect to the offsprings facing direction. The other unit encodes
the distance between parent and offspring (also mapped in the interval between 0 and 1).
The remaining two input units are set to a fixed value of 0. When the individual reaches
adulthood, the first two units are set to 0 and the other two units encode the angle and
distance of the nearest piece of food. The two output units, in both young and adult,
encode four possible motor actions: go one step forward (11), turn 90 left (10) or right
(01), do nothing (00).
Although both young and adults move in the environment on the basis of
environmental sensory information, they are informed by their senses about different
aspects of the environment. Young individuals are in-formed about the current position of
their parent. Adults are informed about the position of the currently nearest piece of food.
Young and adult individuals must use this environmental information in different ways.
The objective of young individuals is to receive food from their parents because this is
the only way for them to survive. Their parents have part of the responsibility because if
they are unable to capture food or if they decide not to give their food to their offspring,
the offspring will not eat. However, another part of the responsibility for eating lies with
the young themselves. The reason is that a young individual can receive food from its
parent only if it is located sufficiently close to its parent. Otherwise, no giving of food
from parent to offspring can take place. This implies that young individuals must use the
environmental (sensory) information about the current position of their parent to generate
motor actions that keep the distance between them and their parent within the required
radius. Since their parent in the meantime is moving itself in the environment in order to
find food, they must follow their parent.
For adults the situation is different. Adults must use environmental (sensory)
information about the location of food to move in the environment so as to reach the food
in as few cycles as possible. (When reached, a food element disappears from the
environment.) Then they must decide whether to use the energy resulting from the
captured food for themselves or to give the food (energy) to their offspring, assuming the
offspring is at the appropriate distance. (If an adult has more than a single young
offspring, it is the nearest offspring which receives the food.)
Artificial societies
206
What decides whether a parent will act egoistically or altruistically towards its
offspring? We assume that each individual inherits from its parent not only the
connection weights of its network (these weights determine how good an individual is at
following its parent when young and at capturing food when adult) but also a special
parameter (gene) which determines the probability of an individual acting egoistically
or altruistically as an adult. Hence, different networks are born as more or less altruistic
individuals. If the inherited value of the altruism gene is high, when the individual
becomes a parent it will tend to behave altruistically and give its energy to its offspring.
If the inherited value is low, the individual will tend to act egoistically and consume all
the energy derived from the captured food.
It is clear that an intermediate, optimal, value of the gene must be found if the
population is not to become extinct. If the average altruism of the population is too
pronounced, each individual will generate very few offspring and for this reason the
population will be at risk of becoming extinct. If the population tends on average to act
too egoistically, the average individual will generate many offspring but these offspring
will tend to die (of starvation) before reaching adulthood and reproducing; in this case too
the population runs the risk of extinction. An intermediate level of altruism means that
the average adult will eat enough to survive and generate new offspring and, at the same
time, will give enough energy to its existing offspring to allow them to survive to
adulthood and reproduce.
In the simulations we are describing, the appropriate level of altruism evolves
spontaneously in the population as does any other trait and does not need to be
determined a priori by the researcher. Any trait can evolve in a population if (a) it varies
from one individual to another; (b) it is inherited; and (c) it is subject to random
mutations. All three properties characterize both the weight matrices and the value of the
altruism gene. Hence, we can observe how the appropriate level of altruism evolves
spontaneously in the population. Figure 12.2 shows how the average tendency to behave
altruistically changes across the generations.
Altruism starts at around 0.5 due to the random assignment of values (between 0 and
1) to the altruism gene of individuals of the initial population and it quickly increases
to stabilize at a value between 0.8 and 0.9.
The evolved behaviour which is of particular interest here, however, is the tendency of
young individuals to follow their parents during the latters movements in search of food.
In fact, in these simulations it is possible to observe the emergence of a type of behaviour
which tends to keep different
207
Artificial societies
208
209
Artificial societies
210
social group together. One might expect this second ability to evolve naturally as a byproduct of the present evolutionary scenario, since by keeping close to their offspring
parents can increase the likelihood that their offspring will receive the energy that allows
them to survive to adulthood and reproduce.
One problem which might arise in the new simulations is that a parent might find itself
in situations in which it has to move in one direction to capture food and in a different
direction to keep close to its offspring. Hence, in such circumstances the parent would
need to decide between two conflicting courses of action. However, as Floreano (1993)
has shown, this problem can be solved by neural networks similar to ours with no need,
again, for explicit rules to solve the conflict. In his simulations organisms live in an
environment which contains a nest in a particular location and a number of randomly
distributed pieces of food. Fitness is only increased when an individual first captures a
(single) piece of food and then goes to the nest. In other words, it is assumed that food
must be captured in the open but can only be consumed in the nest. Floreanos networks
are able to evolve this type of alternating behaviour (food, nest, food, nest, etc.) provided
they are informed by a dedicated set of input units whether they have a piece of food in
their hand or not. In the same way we can expect that, rather than being caught in
unresolvable doubt about what to do in a situation of conflict, our adult networks would
evolve some sort of alternating behaviour of first capturing food and then taking the food
to the offspring.
The social groups of the present simulations are very small. In fact, they are limited to
one parent and its offspring. Two kinds of extension can easily be envisaged, one in the
genetic kin line and one in the non-genetic (affinal) kin line. The extension in the genetic
kin line requires that more than two successive generations live together and form social
groups (families). In our simulations it sometimes happens that some offspring of a living
adult reach adulthood and reproduce. Hence, individuals belonging to three successive
generations may be living at the same time: offspring, parent and grandparent. However,
there can be no tendency for individuals belonging to three successive generations to live
together as a social group because this would require that not only pre-adult offspring
follow their parents but also that adult offspring do the same with their parents. This is
not possible in the present simulations because when an individual reaches adulthood it
ceases to perceive its parent and starts perceiving food. Hence, adult offspring cannot
follow their invisible parents.
However, we might easily allow individuals to continue to perceive their parents after
reaching adulthood while at the same time perceiving and looking for food. The real
problem is what adaptive function remaining near its parent would have for an adult.
Adults obtain their energy by capturing food, not by being given energy by their parents.
Hence, there would be no reproductive advantage in following ones parent after reaching
adulthood. Only by introducing some such advantage might one observe the evolution of
larger social groups constituted by individuals belonging to three, rather than two
successive generations. (Another possibility, of course, is that juvenile traits are
maintained in adulthood even if they are no longer functional.)
A possible reproductive advantage for adult (or even pre-adult) individuals of staying
close to ones parent is that proximity might allow individuals to learn from their parents.
This raises the whole issue of cultural transmission of behaviour through imitation or,
more generally, through learning from others. Imitation can be modelled as a network
211
(the imitator) taking the output of another network (the model) in response to a shared
input as its teaching input.1 In fact, in the simulations described in Denaro & Parisi
(1994) we have shown that neural networks evolve a genetically inherited tendency to
stay close to other networks (independent of any genetic relationship among them) if this
allows them to learn by imitation some useful behaviour that increases their fitness.
An increase in the size of family groups could also be caused by the emergence of
altruistic behaviour with respect to other kin, e.g. cousins. Kin selection theory predicts
that altruistic behaviour is a function of degree of genetic relatedness. Of course, to act
altruistically towards cousins (even if less altruistically than towards offspring or sisters)
requires the parallel development of an ability to recognize different types of kin (Hepper
1991).
The other extension of social groups that might be considered is in the non-genetic kin
line. This extension requires that reproduction ceases to be agamic, as in our simulations,
and becomes sexual, with two parents giving birth to offspring. (For sexual reproduction
in ecological neural networks, see Menczer & Parisi 1992.) Sexual reproduction can have
consequences for the formation of social groups at two levels. First, sexual reproduction
requires that two individuals come together in space for mating to occur, hence sexually
reproducing networks might evolve a tendency to form (temporary) social groups
composed of two individuals. (Werner & Dyer (1991) have simulated communication
between male and female networks that causes two individuals to come together for
mating.) Second, sexual reproduction might cause the creation of more permanent social
groups if the two parents have to live close together in order to help their offspring
survive to adulthood. The two parents might live together as a by-product of living near
the same offspring, or they might live together for some specific function connected with
their offspring. For example, one parent might evolve a tendency to feed not only its
offspring but also its mate. This behaviour might evolve if the survival of the mate is
critical for the survival of the offspring. (The other parent might protect the offspring
from danger.) Hence an altruistic behaviour (feeding) toward a non-genetically related
individual (ones mate) might evolve because it has an indirectly favourable effect on the
reproductive chances of ones offspring. The net result would be the emergence of
permanent social groups consisting of father, mother and offspring.
If the emerging family groups can be enlarged in both the consanguineal and affinal
lines (Murdock 1949), the number of individuals that constitute a social group can
increase considerably. Dunbar (1993) has estimated a group size of around 150
individuals for early social humans, based on an extrapolation of the observed
relationship between group size and neo-cortex size in primates. Ethnographic evidence
on hunter-gatherer and other early types of society indicates a layered group structure
with three levels: bands or overnight groups (3050 individuals); clans and lineage
groups (100200 individuals); and tribes (5002500 individuals) (cf. Dunbar 1993).
Johnson & Earle (1987) distinguish between family-level groups (3050 individuals) and
local groups (300500 individuals).
Of course, the next step in the evolution of human social groups is the emergence of
groups of individuals living together when the individuals which constitute the group are
not relatives. While it may be reasonable to evolve social groups constituted biologically
by kin, as in our simulations, since the basis of the behaviour that leads to the formation
and maintenance of kin groups appears to be mainly genetic, for non-kin social groups it
Artificial societies
212
is more likely that the behaviour which creates such groups is learned from others and,
therefore, culturally rather than genetically transmitted. The general tendency to learn
from others can be inherited biologically in organismssuch as humansthat have
developed an adaptation which assigns a critical role to cultural transmission, but the
actual behaviours which create non-kin social groups are likely to be learned rather than
inherited genetically.
Notes
1. We are assuming back-propagation (Rumelhart, McClelland, the PDP Group 1986) as the
learning method here. For work using this neural network model of imitation see Chapter 9
and Terna 1993.
Chapter 13
Understanding the functions of norms in
social groups through simulation
Rosaria Conte and Cristiano Castelfranchi
In the Artificial Intelligence (AI) literature a norm is treated as a behavioural constraint;
that is, a reduction of the action repertoire and therefore of the actions physically
available to the system. According to this view, norms allow co-ordination within social
systems, the latter being defined as groups of jointly acting agents (Shoham &
Tenneholtz 1992a, b, Moses &Tenneholtz 1992).
This view is insufficient because it accounts exclusively for what are known as norms
of co-ordination (Conte & Castelfranchi 1993). Within game theory, norms of coordination are seen essentially as conventions; that is, behavioural conformities that do
not presuppose explicit agreements among agents and emerge from their individual
interests (Schelling 1960, Lewis 1969). The function of these norms is essentially that of
permitting or improving co-ordination among participants. A well-known example
(Lewis 1969:5) is that of two people unexpectedly cut off during a telephone
conversation. They both want to continue the conversation and each of them has to
choose whether to call back or wait. A convention gradually emerges from interactional
practice, establishing who should do what. Norms of co-ordination, therefore, do not stem
from a conflict of utility. Indeed, the solution of a problem of co-ordination is such that,
if one agent succeeds, so does the other. The single agents utility implies the joint one,
and vice versa.
However, we claim that norms of co-ordination are only a subset of norms. First,
norms are not necessarily conventional: they do not necessarily emerge from implicit
agreement but are also imposed by explicit prescriptions, directives and commands.
Secondly, norms often arise from conflicts of utilities among agents, and sometimes they
provide what economists call Pareto-inferior solutions; that is, solutions which produce a
state of affairs such that some agents are worse off than they were before. Hence, while
norms of co-ordination provide unanimous solutions, this is not true for every type of
norm. Thirdly, not all norms serve to improve co-ordination, reduce accidental collisions
and other negative interferences. Some important norms have other socially desirable
effects. For example, the norm of reciprocation is essentially orientated to control
cheating. Other norms prevent more open forms of deliberate aggression (physical
attack). A good theory of norms should stay distinct and explore their relative
functionalities.
Elsewhere (Conte & Castelfranchi 1993), we have tried to account for norms as
prescriptions represented in the minds of autonomous (that is, self-interested) agents in a
multi-agent context. Here, instead, we are concerned with the functions of norms,
independent of the agents makeup. One specific function will be hypothesized to lie in
Artificial societies
214
the control and reduction of aggression among agents in a common world; that is, in a
world where ones actions have consequences for anothers goal achievements.
215
1 times the number of steps for moving, and 4 for attacking) are also kept constant
throughout the simulation. Of course, the agents strength does change over the course of
each game as a function of the amount of food eaten, the actions performed, and the
attacks suffered by each agent (minus 4 for each attack, no matter who is the winner).
The routines
At present, 50 agents are included in each match. They move in search of food as soon as
some food item appears within their horizon (extending two steps in each direction
from the agents current location). The agents move as castles on a chess board: one step
either up, down, left or right, but not diagonally. Food may be seen within a crossshaped area called a territory, consisting of the four cells to which an agent can move
in one step from its current location). Food can also be smelled within each agents
horizon. At the corners and at the edges of the grid, both territories and horizons are
reduced. Even if a cell in some agents territory is occupied by another agent the latter
does not block the formers sensing of the food items (for example, in Figure 13.1, where
only a region of the entire grid is shown, the food item located at cell (4; 5) may be smelt
by F).
Artificial societies
216
(b) Choose action. This function is called by the food search. The agenda is checked and
if more than one alternative is found, one is chosen at random.
217
Pre-condition:
x is one step from the destination.
Effect:
The order of preference among these actions is as follows: EAT, MOVE-TO-FOODSEEN, MOVE-TO-FOOD-SMELT, ATTACK, MOVE-RANDOM and STAY.
Artificial societies
218
Some examples
To give a flavour of the simulation, let us consider the following three examples (see
Figure 13.2):
219
Artificial societies
220
221
Artificial societies
222
223
In sum, the control of aggression per se is neutral: it is neither pro-social, nor antisocial. Whether it plays a positive or a negative social role depends on the type of control
employed. In the Strategic condition, all the social costs of controlling aggression are
sustained by the weaker individuals. The stronger ones do not pay anything. The stronger
agents are never attacked and therefore do not pay the costs of being attacked, while the
weaker ones do. Stronger agents only pay the costs of undertaking advantageous
aggression. Their costs are lower than their gains (cost of aggression=4, against the food
items nutritional value=20). The weaker agents sustain the costs of being attacked
without obtaining anything. In other words, advantageous aggression is not constrained,
while disadvantageous aggression is.
In the other conditions, although aggression is always to the advantage of the stronger
agents and to the disadvantage of the weaker ones, the costs of being attacked are shared
equally by both weak and strong. Furthermore, in the non-Strategic conditions,
unintentional alliances among weaker agents are allowed which might prove
advantageous for one (as in the example shown in Figure 13.2) or for all of them (as
when a stronger eater attacked by a number of players dies during the match). An indirect
effect of multiple aggression is thus the reduction of social differences.
Even the Blind condition is more equitable than the Strategic one: it allows for
disadvantageous aggression, and although this is a short-term cost for weaker agents, it
turns out to be a long-term benefit as well, in that it reduces the social gap.
It is not surprising that in the Normative condition the distribution is more equitable.
The costs of the control of aggression are equally shared among the strong and the weak.
Because food is constant and the match length is fixed, the differences between agents
introduced by the norm of precedence are not maintained throughout the match. Every
time a food item reappears, the status of some agents changes. Some agents, who earlier
may have been non-possessors, now find themselves endowed with a new possession,
while some others, who have been privileged in the past, become members of the unlucky
part of the population. Thus, the norm in question is not truly a norm of property but
rather a finder-keeper norm, since the possessions are not ascribed exclusively on the
grounds of the right of birth, but may be acquired over the course of each agents life.
The norm of finder-keeper helps the agents to find a solution to competition for the same
food item, if any such competition occurs. Interestingly, this type of norm, which seems
to have a biological foundation (de Waal 1982; Eibl-Eibesfeldt 1967/1978), while
controlling aggression efficaciously, also reduces the variance of strength among the
agents; that is, their inequality.
Artificial societies
224
attacked less frequently and consequently eat more food. For the weaker agents, a life of
aggression is not convenient. A distributive view of norms is suggested initially by our
findings: some social laws are useful for the weaker to control the dominance of the
stronger.
A problem arises immediately from this conclusion: why should the stronger obey a
norm which tends to favour the weaker more than it favours them? Why should all agents
accept a non-unanimous sort of solution to social conflicts? These are classical questions
of moral and political philosophy that we cannot solve merely on the grounds of our
simplistic experiments. However, let us at least try to formulate some answers to the
questions:
(a) Additional costs in norm violation. In simulating the prescriptive character of norms
among cognitive agents, a spreading awareness of the norm, the control behaviour,
and short-term and long-term punishments would be expected. Under these conditions,
cheats could receive high additional costs (punishment for transgression), and the
stronger are more likely than the weaker to cheat. The existence of the norm does not
favour the stronger (indeed, it is a pure cost), but once the norm has been established,
to observe it is advantageous, even for the stronger.
(b) Common interest. Whether or not the population of agents considered does really
form a group makes a large difference. Members do not pursue only individual
interests, but also some common goals of mutual advantage. In such cases, the control
of intra-group aggression becomes crucial. We may imagine three different scenarios:
(i) Inter-group competition. Suppose there is a prize for the group that obtains the highest
strength score, to be distributed among its members. In this situation, it might be better
for the stronger agents to cooperate by observing the norm and thus reducing the
level of aggression.
(ii) Need for mutual aid and exchange. Suppose that the members of the population need
to help one another to realize their individual goals (i.e. there are dependence
relations: Castelfranchi et al. 1992; Conte et al. 1991, Cesta & Miceli 1993), and
suppose that agents are able and willing to provide such help only when they have a
certain amount of power. In this case, the stronger agents may have some interest in a
relative decrease of their personal strength and in a more equitable distribution.
(iii) The wheel of fortune. A third possible explanation points to different degrees of
variability in the agents fates. In the wheel of fortune, you cannot tell who will climb
and who will sink on the social scale. Even in our artificial world, the status of the
agents during the game is not constant: powerful agents can be reduced to weak, and
vice versa. This is due to fortune (and misfortune) and to the effects of attacks
(especially multiple attacks in which the aggressors pay just a share of the costs
sustained by the target). If the stronger can lose their status and power, it is possible to
ascertain to what extent and in which condition it is better for them to respect norms.
For example, a norm of reciprocity might be better for the stronger to observe (be
merciful and you will receive the same treatment if and when you become weaker).
However, the question of whether the agents fates are steadier in normative scenarios
than in the non-normative conditions is an open one. In our normative condition, fates
are at least in part precarious because possessions change over the course of the match.
However, it would be interesting to try out this hypothesis with other types of norm.
225
Concluding remarks
Experimental studies of the emergence of co-operation (e.g. Kreps et al. 1982, Axelrod
1984, Axelrod & Dion 1988, Lomborg in press, Bicchieri 1990, Glance & Huberman
1993, Huberman & Glance 1993) have achieved a level of sophistication which is far
beyond that of the experiment presented here. Unlike those studies, however, this
experiment is meant to pave the way for the study of norm functions among cognitive
agents, and, consequently, to introduce both mental representations of norms and
normative reasoning in the cognitive agents. Therefore, the most interesting findings are
expected to be found through developing more sophisticated agent architectures, rather
than experimenting with larger populations, more sophisticated dynamics, or more
complex marginal utility functions.
The present experiment seems to show that:
(a) norms may have the function of constraining aggression;
(b) by controlling aggression, norms also have an indirect positive effect on the strength
of the population (at least as effective as utilitarian constraints);
(c) norms have a specific and positive impact on the single agents share of the overall
cake. This seems to be a result of the specific type of aggression control allowed by
normative constraints, which reduce both advantageous and disadvantageous attacks,
thus redistributing the costs of controlling aggression over the population. In the nonNormative conditions, advantageous attacks were not reduced and the costs of
aggression control were not fairly distributed.
Future work will aim to:
(a) implement additional types of norm (e.g. the norm of reciprocity) and examine their
consequences;
(b) compare different strategies of aggression control within the same population, and
observe their evolutionary trends;
(c) allow some learning mechanism, and observe the spreading of normative strategies;
and
(d) allow some level of normative choice by implementing a cognitive model of norms.
Acknowledgements
We would like to thank Juri Castelfranchi and Vincenzo Catoni for their help in
programming, Alessandro Laudanna and Dario Salmaso for their competent (and patient)
statistical suggestions, and Fabrizio Stasolla for his assiduous and understanding
collaboration.
Notes
1. In an earlier stage of the experiment, two norms of precedence were applied, namely a norm
of precedence based upon possession of resources, and a convention forcing agents to give
precedence to eaters to their right. Both fitted the simulated environment rather well, but
while the former was grounded on natural and historical conditions, the latter was
Artificial societies
226
merely formal and contractual. In that stage of the experiment, however, the simulations
were run without the food supply being kept constant. As a result, the initial possession of
food items introduced a difference among the players that was maintained throughout the
game. In contrast, at the present stage, the food supply is kept constant, and the first norm
has been changed to a finder-keeper norm: precedence is given to the agents in whose
territory a given food item happens to fall. The difference between the first and second
norms is much less evident. Therefore, at the present stage, only one type of normative
condition, namely precedence to possessors, is used.
Chapter 14
A logical approach to simulating societies
Michael Fisher and Michael Wooldridge
A number of languages and testbeds for implementing and simulating artificial social
systems have been described in the literature of Distributed Artificial Intelligence (DAI)
and its related disciplines (see, for example, Hewitt 1977, Agha 1986, Gasser et al. 1987,
Shoham 1990, Bouron et al. 1991, Doran et al. 1991, Ferber & Carle 1991). However,
with the possible exception of pure Actor languages (Hewitt 1977, Agha 1986), these
tools have relied on an informal, if not ad hoc, notion of agency, with similarly informal
techniques for programming agents. As a result, it is difficult to reason, either formally or
informally, about the expected behaviour of systems implemented using such tools. It is
similarly difficult to derive an a posteriori explanation of such a systems behaviour. In
this chapter, we present a language for simulating co-operative and competitive
behaviour in groups of intelligent artificial agents. In contrast to the informal languages
and testbeds referred to above, this language, called Concurrent METATEM, has a
simple, well-motivated logical (and hence mathematical) foundation. In relative terms, it
is easy to reason about Concurrent METATEM systems at both formal and informal
levels (Fisher and Wooldridge 1993). Concurrent METATEM is, so far as we are aware,
unique in DAI in having such a well developed and motivated logical foundation.
This chapter is structured as follows. In the remainder of this section, we motivate the
language and discuss its origins in more detail. In the next section, we present a more
detailed, though essentially non-technical, introduction to the language and its execution.
We then demonstrate by example how the language may be used to model societies.
Finally, we describe the current state of the language and describe our intentions with
respect to future developments.
Distributed AI is a relatively young discipline which has developed its own tools for
building, experimenting, and evaluating theories and applications. Over the past decade,
many frameworks for building DAI systems have been reported. While some, such as the
DVMT, allow experimentation only with one specific scenario (Durfee 1988) and others
provide extensions to existing AI languages such as LISP (Gasser et al. 1987),
comparatively few fundamentally new languages have been proposed for DAI. Arguably
the most successful DAI languages have been based, at least in part, on the Actor model
of computation (Hewitt 1977, Agha 1986); for example, Ferber and Carle (1991) and
Bouron et al. (1991).
It is our contention that while frameworks for DAI based on extensions to existing AI
languages are useful for experimentation, they will not, ultimately, be viable tools for
building production DAI systems. This is because DAI systems are a subset of the class
of systems known in mainstream computer science as reactive systems. A reactive
system in this sense is one whose purpose is to maintain some ongoing interaction with
its environment. We therefore do not use the term reactive system in the way that it has
Artificial societies
228
become fashionable to use it in AI, to refer to systems that respond directly to their
environment, perhaps using situation-action rules, without reasoning in any way. All
concurrent or distributed systems are reactive in the sense we intend, just as a module or
agent in a concurrent system must maintain some interaction with the other components
in the system (see, for example, Pnueli (1986) for a discussion of this point). It is for this
reason we claim that all DAI systems are reactive.
Contemporary DAI testbeds and languages are usually built as extensions to the
classic AI languages LISP and PROLOG. However, neither of these languages is well
suited to building reactive systems. They are both based on a quite different view of
systems, called the functional or relational view. For this reason, we suggest that an
obvious development for DAI is to build languages based on a more reactive view;
Concurrent METATEM is such a language.
In a 1977 paper, Pnueli proposed temporal logic as a tool for reasoning about reactive
systems. When describing a reactive system, we often wish to express properties such as
if a request is sent, then a response is eventually given and such properties are easily
and elegantly expressed in temporal logic. However, proving the properties of reactive
systems using temporal logic is not a trivial matter. The difficulties involved in this
process led, in the early 1980s, to the idea of using temporal logic itself as a
programming language (Moszkowski 1986). This idea led directly to the METATEM
concept (Barringer et al. 1989). METATEM is a general framework for executing
temporal logic specifications, where these specifications are expressed as a set of
temporal logic rules. The concept of a reactive system is therefore at the very heart of
METATEM. Although the original METATEM proposal did not address the issue of
concurrency, the potential value of concurrently executing METATEM systems,
particularly for DAI applications, was recognized immediately (Fisher and Barringer
1991). Concurrent METATEM is a simple operational framework that allows societies of
METATEM processes to communicate and co-operate.
Note that although Concurrent METATEM may be regarded as a logic programming
language, in that it has a well-developed and motivated logical foundation, it is quite
unlike any other logic programming language with which we are familiar; in particular, it
is based on a novel model for concurrency in executable logic. Whereas most previous
concurrent logic paradigms are based on fine-grained AND-OR parallelism, (e.g. Clark
and Gregory 1987), concurrency in Concurrent METATEM is achieved via coarsegrained computational entities called agents; each agent is a METATEM process.
Concurrent METATEM
In Concurrent METATEM (Fisher 1993), the behaviour of an agent is defined using a
temporal logic formula. Temporal logic is used because, not only is it an ideal formalism
in which to represent the dynamic properties of an agent, but it also contains an explicit
mechanism for representing and executing goals. As well as providing a declarative
description of the agents desired behaviour, the temporal formula can be executed
directly to implement the agent (Fisher and Owens 1992). Thus, the basic behaviour of an
agent consists of following a set of temporal rules representing the basic dynamics of the
agent, introducing new goals, and attempting to satisfy existing goals.
229
Artificial societies
230
Here, {pop, push} is the set of environment predicates, (i.e. messages the agent
recognises), while {popped, stackfull} is the set of component predicates (i.e. messages
the agent itself might produce). Note that these sets need not be disjointedan agent may
broadcast messages that it also recognizes. In this case, messages sent by an agent to
itself are recognized immediately.
During the execution of each agents rules, the two types of predicate have a specific
operational interpretation, as follows:
(a) Environment predicates represent incoming messages. An environment predicate can
be made true if, and only if, the corresponding message has just been received. Thus, a
formula containing an environment predicate, such as request(x), where x is a
231
universally quantified variable, is only satisfied if a message of the form request (b)
has just been received (for some argument b).
(b) Component predicates represent messages broadcast from the agent. When a
component predicate is satisfied, it has the (side-)effect of broadcasting the
corresponding message to the environment. For example, if the formula offer(e) is
satisfied, where offer is a component predicate, then the message offer (e) is broadcast.
Note that some predicates used by an agent are neither environment nor component
predicates; these internal predicates have no external effect. They are used as part of the
internal computation of the agent and, as such, do not correspond either to message
sending or message reception.
Once an agent has commenced execution, it follows a cycle of reading incoming
messages continuously, collecting together the rules that fire (i.e. whose left-hand sides
are satisfied by the current history) and executing one of the disjuncts represented by the
conjunction of right-hand sides of fired rules. Individual agents execute
asynchronously, and are autonomous in that they may execute independently of incoming
messages and may change their interface dynamically.
Agents may backtrack, with the proviso that an agent may not backtrack past the
broadcasting of a message. Consequently, in broadcasting a message to its environment,
an agent effectively commits the execution to that particular path. Thus the basic cycle of
operation for an agent can be thought of as a period of internal execution, possibly
involving backtracking, followed by appropriate broadcasts to its environment.
Finally, agents are also members of groups. Each agent may be a member of several
groups. When an agent sends a message, that message is, by default, broadcast to all the
members of its group(s), but to no other agents. Alternatively, an agent can choose to
broadcast only to certain of the groups of which it is a member. This mechanism allows
the development of complex structuring within the agent space and provides the potential
for innovative applications, such as the use of groups to represent physical properties of
agents. For example, if we assume that any two agents in the same group can see each
other, then movement broadcast from one agent can be detected by the other. Similarly, if
an agent moves out of sight, it moves out of the group and thus the agents that remain
in the group lose sight of it. Examples such as this (which will not be explored further
here) show some of the power of the group concept.
Examples
In this section we shall give a range of examples, starting with collections of agents
acting individually and adding gradually structure and the potential for more complex
interactions amongst agents and groups of agents. We will begin with a simple scenario,
taken from Fisher (1993), where seven agents are each attempting to get a resource from
a single agent.
Artificial societies
232
To give some idea of how Concurrent METATEM can be used, a simple example
showing agents with a form of intelligent behaviour will be described. First, a brief
outline of the properties of the leading characters in this example will be given.
The scenario
Snow White has a bag of sweets. All the dwarfs want sweets, though some want them
more than others. If a dwarf asks Snow White for a sweet, she will give him what he is
asking for, but perhaps not straight away Snow White is only able to give away one sweet
at a time.
Snow White and the dwarfs are going to be represented as a set of agents in
Concurrent METATEM. Each dwarf has a particular strategy that it uses in asking for
sweets, as described below (the names and behaviours of the dwarfs differ from those in
the fairy tale!).
1. eager initially asks for a sweet and, from then on, whenever he receives a sweet, asks
for another.
2. mimic asks for a sweet whenever he sees eager asking for one.
3. jealous asks for a sweet whenever he sees eager receiving one.
4. insistent asks for a sweet as often as he can.
5. courteous asks for a sweet only when eager, mimic, jealous and insistent have all
asked for one.
6. generous asks for a sweet only when eager, mimic, jealous, insistent and courteous
have all received one.
7. shy only asks for a sweet when he sees no one else asking.
8. snow white can only allocate one sweet at a time. She keeps a list of outstanding
requests and attempts to satisfy the oldest one first. If a new request is received, and it
does not occur in the list, it is added to the end. If it does already occur in the list, it is
ignored. Thus, if a dwarf asks for a sweet n times, he will eventually receive at most n,
and at least 1, sweets.
This example may seem trivial, but it represents a set of agents exhibiting different
behaviours, where an individual agents internal rules can consist of both safety (e.g.
nothing bad will happen) and liveness (e.g. something good will happen) constraints,
and where complex interaction can occur between autonomous agents.
The program
The Concurrent METATEM program for the scenario described above consists of the
definitions of eight agents shown below. To provide a better idea of the meaning of the
temporal formulae representing the internals of these agents, a brief description will be
given with each agents definition. Requests to Snow White are in the form of an ask ( )
message with the name of the requesting dwarf as an argument. Snow White gives a
sweet to a particular dwarf by sending a give ( ) message with the name of the dwarf as
an argument. Finally, uppercase alphabetic characters, such as X and , represent
universally quantified variables.
233
1.
Initially, eager asks for a sweet and, whenever he has just received a sweet, he asks again.
2.
If eager has just asked for a sweet then mimic asks for one.
3.
If eager has just received a sweet then jealous asks for one.
4.
From the beginning of time, insistent asks for a sweet as often as he can.
5.
If courteous has not asked for a sweet since eager asked for one; has not asked for a sweet since
mimic asked for one; has not asked for a sweet since jealous asked for one; and has not asked
for a sweet since insistent asked for one, then he will ask for a sweet.
6.
If generous has not asked for a sweet since eager received one; has not asked for a sweet since
mimic received one; has not asked for a sweet since jealous received one; has not asked for a
sweet since insistent received one; and has not asked for a sweet since courteous received one,
then he will ask for a sweet!
7.
Initially, shy wants to ask for a sweet but is prevented from doing so whenever he sees some
other dwarf asking for one. Thus, he only succeeds in asking for a sweet when he sees no one
else asking and, as soon as he has asked for a sweet, he wants to try to ask again.
Artificial societies
234
8.
If snow-white has just received a request from a dwarf, a sweet will be sent to that dwarf
eventually. The second rule ensures that sweets can not be sent to two dwarfs at the same time,
by stating that if both give (X) and give (Y) are to be broadcast, then X must be equal to Y.
Note that, in this example, several of the dwarfs were only able to behave as required
because they could observe all the ask ( ) and give ( ) messages that were broadcast. The
dwarfs can thus be programmed to have strategies that are dependent on the behaviour of
other dwarfs. Also, the power of executable temporal logic is exploited in the definition
of several agents, particularly those using the operator to represent multiple goals.
We also note that, as the agents behaviour is represented explicitly and in a logical
way, verification of properties of the system is possible. For example, given the agents
definitions, we are able to prove that every dwarf except shy will eventually receive a
sweet. For further work on the verification of properties of such systems, see Fisher and
Wooldridge (1993).
Adding money
We now extend this example to incorporate the notion of some resource which the dwarfs
can attempt to exchange with Snow White for sweets, i.e. money.
Bidding
Initially, we will simply change the ask predicate so that it takes an extra argument
representing the amount the dwarf is willing to pay for a sweet. This enables dwarfs to
bid for a sweet, rather than just asking for one. For example, dwarf1 below asks for a
sweet, bidding 2.
dwarf1( ) [ask] :
start
ask (dwarf1, 2)
We can further modify a dwarfs behaviour so that it does not bid more than it can afford,
by introducing some record of the amount of money that the dwarf has at any one time.
Thus, the main rule defining the bidding behaviour of a dwarf might become something
like
Note that the behaviour of Snow White might also change so that all the bids are recorded
and then a decision about which bid to accept is made based upon the bids received. Once
a decision is made, give ( ) is again broadcast, but this time having an extra argument
showing the amount paid for the sweet. For example, if Snow White accepts the bid of
2from dwarf1, then give (dwarf 1, 2) is broadcast.
235
Finally, a dwarf whose bid has been accepted, in this case dwarf1, must remember to
record the change in finances:
Renewable resources
Dwarfs who keep buying sweets will eventually run out of money. Thus, we may want to
add the concept of the renewal of resources, i.e. being paid. This can either happen
regularly, at time intervals defined within each dwarfs rules, for example:
Or the dwarf can replenish its resources when it receives a particular message from its
environment, for example:
dwarf1 (go) [ask]:
Competitive bidding
As the bids that individual dwarfs make are broadcast, other dwarfs can observe the
bidding activity and revise their bids accordingly. We saw earlier that the mimic dwarf
asks for a sweet when it sees the eager dwarf asking for one. Similarly, dwarf2 might
watch for any bids by dwarf1 and then bid more, for example:
Although we will not give further detailed examples in this vein, it is clear that a range of
complex behaviours based upon observing others bids can be defined.
Co-operation
In the previous section we showed how individual dwarfs might compete with each other
for Snow Whites sweets. Here, we will consider how dwarfs might co-operate in order
to get sweets from Snow White. In particular, we consider the scenario where one dwarf
on its own does not have enough money to buy a sweet and thus requires a loan from
other dwarfs.
Borrowing money
In order to borrow money from other dwarfs to buy a sweet, a dwarf can broadcast a
request for a certain amount. For example, if the dwarf (dwarf3 in this case) knows that
the highest amount bid for a sweet so far is X and he only has , he can ask to borrow X
, possibly as follows.
Artificial societies
236
Now, if another dwarf, say dwarf4, offers to lend a certain amount, say Z, to dwarf3, then
another rule recording the loan must be added to dwarf3s rule-set:
Lending behaviour
Dwarfs might have various strategies for lending and borrowing money. For example,
perhaps a dwarf will not lend any more money to any dwarf who still owes money.
Further, a dwarf might be less likely to lend money to any dwarf who has never offered to
help his previous requests. Again, a variety of strategies for lending and borrowing can be
coded in Concurrent METATEM. Rather than giving further examples of this type, we
next consider the use of groups in the development of structured systems of interacting
agents.
Group structuring
As described earlier, as well as the notion of autonomous objects, Concurrent
METATEM also provides a larger structuring mechanism through groups. This
restricts the extent of an objects communications and thus provides an extra mechanism
for the development of strategies for organizations. Rather than giving detailed examples,
we will outline how the group mechanism could be used in Concurrent METATEM to
develop further co-operation, competition and interaction between agents.
Again, we will consider a scenario similar to that of Snow White and the Seven
Dwarfs, but will assume the existence of a large number of dwarfs and possibly several
Snow Whites. We will outline several examples of how the grouping of these agents can
be used to represent more com- plex or refined behaviour.
Collective bidding
If we have a situation where dwarfs bid for sweets, we can organize co-operation within
groups so that the group as a whole puts together a bid for a sweet. If successful, the
group must also decide to whom to distribute the sweet. Thus a number of groups might
be co-operating internally to generate bids, but competing (with other groups) to have
their bid accepted.
Forming subgroups
Within a given group, various subgroups may be formed. For example, if several
members of a group are unhappy with another members behaviour, they might be able to
create a new subgroup which excludes the unwanted agent within the old grouping. Note
that members of the subgroup can hear the outer groups communications, while
237
members of the outer one cannot hear the inner groups communications. Although we
have described this as a retributive act, such dynamic restructuring is natural as groups
increase in size. By using a combination of individual agent strategies and of grouping
agents together, we are able to form simple societies. In particular, we can represent
societies where individuals co-operate with their fellow group members, but where the
groups themselves compete for some global resource.
Although our examples have been based upon agents competing and co-operating in
order to obtain a resource, many other types of multi-agent system can be developed in
Concurrent METATEM. It is important to note that there is no explicit global control or
global plan in these examples. Individual agents perform local interactions with each
other and with their environment.
Related work
Concurrent METATEM is in some respects similar to Shohams AGENT0 system
(Shoham 1990). AGENT0 is a first attempt to build an agent oriented programming
(AOP) language. AOP is a new programming paradigm, based on a societal view of
computation (Shoham 1990:4), central to which is the idea of agents/agents as cognitive
entities, whose state is best described in terms of mentalistic notions such as belief,
choice and commitment.
Both AGENT0 and Concurrent METATEM are based on temporal logic, although
these have quite different forms. In Concurrent METATEM, a tense logic approach is
adopted. The language of Concurrent METATEM rules is a classical logic augmented by
a set of modal operators for describing the dynamic nature of the world. In AGENT0, the
method of temporal arguments is used. This means that the language contains terms for
referring directly to time. Predicates and modal operators are then date stamped with
the time at which they were true. Both styles of temporal logic have advantages and
disadvantages and this chapter will not go into the relative merits of each. AGENT0 and
Concurrent METATEM are both rule-based languages, although each makes novel use of
the concept of a rule. In both languages, the rules an agent possesses determine how that
agent makes commitments. In AOP, commitment is given a mentalistic interpretation,
based on the formalisations in Shoham (1989) and Thomas et al. (1991). In contrast,
Concurrent METATEM gives commitment a precise computational meaning. Despite
these similarities, AGENT0 and Concurrent METATEM differ in many significant
respects. An obvious distinguishing feature is the nature of rules in the two languages. In
Concurrent METATEM, rules have an explicit logical semantics and are based on the
separated (past
future) form. In AGENT0, rules do not have such a well-developed
logical semantics.
Another system used extensively for DAI is Georgeff and Lanskys procedural
reasoning system (PRS) (Georgeff & Lansky 1987; Georgeff & Ingrand 1989), which
employs some elements of the belief-desire-intention (BDI) architecture partly
formalized by Rao (Rao and Georgeff 1991). The PRS also has much in common with
Concurrent METATEM: both systems maintain beliefs about the world (in Concurrent
METATEM these beliefs are past-time temporal formulae); knowledge areas in the PRS
loosely resemble Concurrent METATEM rules; and PRS intentions resemble Concurrent
Artificial societies
238
METATEM commitments. There are many points of difference, however. Notably, the
PRS does not have a logical basis for execution that is as elegant as that in Concurrent
METATEM. However, the elegance of Concurrent METATEM does not come cheap.
The PRS seems to be more flexible in execution than Concurrent METATEM and is
likely to have higher potential in time-critical applications.
The computational model underlying Concurrent METATEM is somewhat similar to
that employed in the autonomous agent model of Maruichi et al. (1990). More
generally, there are also some similarities with the Actor model (Hewitt 1977; Agha
1986). The key differences are the ability of agents in Concurrent METATEM to act in a
non-message-driven way and the use of broadcast, rather than point-to-point message
passing.
Finally, a comment on the use of mentalistic terminology. Both AGENT0 and the PRS
make free with terms such as belief, desire, and intention. Shoham argues that such
terminology provides a useful degree of abstraction for complex systems (Shoham 1990,
pp. 56). Although Concurrent METATEM employs terminology such as commitment,
no attempt is made to relate this terminology to a deeper theory of agency (as Shoham
(1990) hopes to do in AOP).
Bibliography
Abraham, R. & C.Shaw 1987. Dynamics: a visual introduction. In Self organizing systems,
F.E.Yates (ed.), 54397. New York: Plenum.
Agha, G. 1986. Actorsa model for concurrent computation in distributed systems. Cambridge,
Mass.: MIT Press.
Alexander, R.D. 1989. Evolution of the human psyche. In The human revolution, P. Mellars &
C.Stringer (eds), 455513.Edinburgh University Press.
Allen, P. 1978. Dynamique des centres urbains. Sciences et Techniques 50, 1519.
Allen, P. & M.Sanglier1979. A dynamic model of growth in a central place system. Geographical
Analysis, 11, 25672.
Arrow, K.J. 1962. The economic implications of learning by doing. Review of Economic Studies
29, 155174.
Arthur, B. 1988. Urban system and historical path dependence. In Cities and their vital systems:
infrastructures, past, present and future, J.H.Ausubel & R.Herman (eds), 8597. Washington,
DC: National Academy Press.
Axelrod, R. 1984. The evolution of cooperation. London: Penguin.
Axelrod, R. & D.Dion 1988. The further evolution of cooperation. Nature 242, 139590.
Axelrod, R. & D.S.Bennett 1993. A landscape theory of aggregation. British journal of Political
Science 23, 21133.
Bairoch, P. 1985. De Jericho Mexico, villes et conomie dans Ihistoire. Paris: Gallimard.
Barringer, H., M.Fisher, D.Gabbay, G.Gough, R.Owens 1989. METATEM: A framework for
programming in temporal logic. In Proceedings of REX Workshop on Stepwise refinement of
distributed systems: models, formalisms, correctness. Mook, Netherlands (published in Lecture
Notes in Computer Science, 430, 94129. Berlin: Springer).
Barringer, H., M.Fisher, D.Gabbay, A.Hunter 1991. Meta-reasoning in executable temporal logic.
In J.Allen, R.Fikes, E.Sandewall (eds), Proceedings of the international conference on
principles of knowledge representation and reasoning (kr), Cambridge, Mass.
Berlekamp, E.R., J.H.Conway, R.K.Guy 1982. Winning ways for your mathematical plays, vol. 2.
New York: Academic Press.
Berry, B.J.L. 1976. Urbanization and Counterurbanization. Los Angeles: Sage.
Bicchieri, C. 1990. Norms of co-operation. Ethics 100, 83861.
Bittner, E. 1974. The concept of organisation. In Ethnomethodology, R.Turner (ed). London:
Penguin.
Boehm, C. 1993. Egalitarian behavior and reverse dominance hierarchy, Current Anthropology 34,
22754.
Bond, A.H. 1991. Commitments and projects. Proceedings of the eighth conference of the society
for the study of artificial intelligence and the simulation of behaviour, 2635, Leeds, UK.
Bond, A.H. & L.Gasser 1988a. An analysis of problems and research in DAI. In Readings in
distributed artificial intelligence, A.H.Bond & L.Gasser (eds), 335, San Mateo, California:
Morgan Kaufmann.
Bond, A.H. & L.Gasser (eds) 1988b. Readings in Distributed Artificial Intelligence. San Mateo,
California: Morgan Kaufmann.
Boudon, R.. 1968. A quoi sert la notion de structure. Paris: Les Essais, Gallimard.
Bouron, T., J.Ferber, F.Samuel1991. MAGES: A multi-agent testbed for heterogeneous agents. In
Decentralized AI 2Proceedings of the second European workshop on modelling autonomous
Bibliography
241
agents and multi-agent worlds (MAAMAW), Y.Demazeau & J.P.Muller (eds), 195214.
Amsterdam: Elsevier/North Holland.
Boyd, R. & P.J.Richerson 1985. Culture and the evolutionary process. Chicago: University of
Chicago Press.
Braudel, F. 1967. Civilisation matrielle et capitalisme. Paris: A.Colin.
Bueno de Mesquita, B. & D.Lalman 1992. War and reason: domestic and international
imperatives. New Haven, Connecticut: Yale University Press.
Buss, L.W. 1987. The evolution of individuality. Princeton, New jersey: Princeton University Press.
Byrne, R. & A.Whiten (eds) 1988. Machiavellian intelligence. New York: Oxford University Press.
Caill, A. 1992. Sujets individuels et sujets collectifs. Philosophie et Anthropologie, coll. Espace
international philosophie, December, 93114. Paris: Les seminaires du Centre Georges
Pompidou.
Camagni, R., L.Diappi, G.Leonardi 1986. Urban growth and decline in a hierarchical system: a
supply-oriented dynamic approach. Regional Science and Urban Economics 16, 14560.
Camerer, C. 1987. Do biases in probability judgement matter in markets? American Economic
Review 77, 98197.
Castelfranchi, C., M.Miceli, A.Cesta 1992. Dependence relations among autonomous agents. In
Decentralized AI3, E.Werner & Y.Demazeau (eds). Amsterdam: Elsevier.
Cesta, A. & M.Miceli 1993. In search of help: strategic social knowledge and plans. Proceedings of
the 12th international workshop on distributed AI, Hidden Valley, Philadelphia.
Champion, A. 1989. Counterurbanization. London: Edward Arnold.
Chattoe, E. 1993. An evolutionary simulation of price and quantity setting with methodological
implications for economics. Unpublished MS, Lady Margaret Hall, Oxford University.
Chauvin, Y. 1988. Symbol acquisition in humans and neural (PDP) networks. Unpublished
doctoral dissertation, University of California, San Diego.
Christaller, W. 1933. Die Zentralen Orte in Sddeutschland. Jena: Fischer.
Cigler, A.J. & B.A.Loomis 1986. Interest group politics (2nd edn). Washington, DC:
Congressional Quarterly Press.
Clancey, W. 1989. Ontological commitments and cognitive models. Proceedings of The eleventh
annual conference of the cognitive science society. Ann Arbor, Michigan.
Clark, E. 1983. Convention and contrast in acquiring the lexicon. In Cognitive development and the
development of word meaning, T.B.Seiler & W.Wannenmacher (eds), 6789. Berlin: Springer.
1987. The principle of contrast: a constraint on language acquisition. In Mechanisms of
language acquisition, B.MacWhinney (ed), 133. Hillsdale, New Jersey: Lawrence Erlbaum.
Clark, K. & S.Gregory 1987. A relational language for parallel programming. In Concurrent
Prolog: collected papers. E.Shapiro (ed.), 926. Cambridge, Mass.: MIT Press.
Coase, R. 1937. The nature of the firm. Economica 4, 386405.
Cohen, M.N. 1985. Prehistoric hunter-gatherers: the meaning of social complexity. In Prehistoric
hunter-gatherers: the emergence of cultural complexity, T.Douglas-Price & J.A. Brown (eds),
99119. New York: Academic Press.
Coleman, J.S. 1993. The rational reconstruction of society. American Sociological Review 58, 1
15.
Collins, R. 1990. Stratification, emotional energy and the transient emotions. In Research agendas
in the sociology of emotions, T.D.Kemper (ed.) 2757. State University of New York Press.
Collins, R.J. & D.R.Jefferson 1991. Ant farm: towards simulated evolution. In Artificial life II,
C.Langton (ed), 579601. Reading, Mass.: Addison-Wesley.
Conte, R., & C.Castelfranchi 1993. Norms as mental objects: from normative beliefs to normative
goals. Proceedings of MAAMAW 93. Neuchatel: Switzerland.
1994. Mind is not enough: precognitive bases of social interaction. In Simulating societies,
N.Gilbert & J.Doran (eds), 26786. London: UCL Press.
1995. Cognitive and social action. London: UCL Press.
Bibliography
242
Conte, R., M.Miceli, C.Castelfranchi 1991. Limits and levels of cooperation: disentangling various
types of prosocial interaction. In Decentralized AI2, Y.Demazeau & J.P.Muller (eds).
Amsterdam: Elsevier.
Conte, R. & J.Sichman. DEPNET: how to benefit from social dependence. Journal of
Mathematical Sociology. Special issue on sociological algorithms, in press.
Corbara, B. 1991. Lorganisation sociale et sa gense chez la fourmi Ectatomma ruidum Roger
(Formicidae, Ponerinae). Doctoral thesis, University of Paris XIII. Grenoble: Microfiche
ANRT/91/PA13/1007.
Corbara, B., D.Fresneau, J-P.Lachaud, Y Leclerc 1991. Evolution de la division du travail dans les
jeunes socits de Neoponera apicalis et dEctatomma ruidum (Formicidae, Ponerinae). Actes
Collectif Insectes Sociaux 7, 18994.
Couclelis, H.1985. Cellular worlds, a framework for modelling micro-macro dynamics.
Environment and Planning A 17, 58596.
1986. Artificial intelligence in geography: conjectures on the shape of things to come. The
Professional Geographer 38, 111.
1988. Of mice and men: what rodent populations can teach us about complex spatial
dynamics. Environment and Planning A 20, 99109.
Courrge, P. 1965. Un modle mathmatique des structures lmentaires de parent. LHomme 5,
24891.
Cusack, T.R. & R.J.Stoll 1990. Exploring realpolitik: probing international relations theory with
computer simulation. Boulder, Colorado: Lynne Rienner.
Davidson, R.H. & W.J.Oleszek 1985. Congress and its members. Washington, DC: Congressional
Quarterly Press.
Davis, R. & R.G.Smith 1983. Negotiation as a metaphor for distributed problem-solving. Artificial
Intelligence 20, 63100.
De Vries, J. 1984. European urbanization, 15001800. London: Methuen.
Denaro, D. & D.Parisi 1994. Imitation and cultural transmission in populations of neural networks.
Submitted to 1994 Conference of the Cognitive Science Society.
Doran, J.E. 1992. A computational investigation of three models of specialization, exchange and
social complexity. In Representations in archaeology, J-C.Gardin & C.S. Peebles (eds), 31528.
Bloomington, Indiana: Indiana University Press.
Doran, J. & N.Gilbert 1994. Social simulation: an introduction. In Simulating societies: the
computer simulation of social phenomena, N.Gilbert & J.Doran (eds), 118. London: UCL
Press.
Doran, J., H.Carvaja, Y.J.Choo, Y.Li 1991. The MCS.multi agent testbed: developments and
experiments. In Proceedings of the international working conference on cooperating knowledge
based systems (CKBS)., S.M.Deen (ed.), 24054. London: Springer.
Doran, J., M.Palmer, N.Gilbert, P.Mellars 1994. The EOS project: modelling Upper Palaeolithic
social change. In Simulating societies, N.Gilbert & J.Doran (eds), 195222. London: UCL
Press.
Dosi, G. & M.Egidi 1991. Substantive and procedural uncertainty: an exploration of economic
behaviours in complex and changing environments. Journal of Evolutionary Economics, 1, 145
68.
Dosi, G., C.Freeman, R.R.Nelson, G.Silverberg, L.Soete (eds) 1988. Technical change and
economic theory. London: Pinter.
Dosi, G., L.Marengo, A.Bassanini, M.Valente (1993). Microbehaviors and dynamical systems:
economic routines as emergent properties of adaptive learning. Berkeley, California: University
of California Press, CCC working paper no. 9313.
Doyle, J. 1992. Rationality and its roles in reasoning. Computational Intelligence 8, 376409.
Drogoul, A., B.Corbara, D.Fresneau 1992a. Applying ethomodelling to social organization in ants.
In Biology and evolution of social insects, J.Billen (ed.), 37583. Leuven University Press.
Bibliography
243
Drogoul, A., J.Ferber, B.Corbara, D.Fresneau 1992b. A behavioural simulation model for the study
of emergent social structures. In Toward a practice of autonomous systems, F.J.Varela &
P.Bourgine (eds), 16170. Cambridge, Mass.: MIT Press.
Duby, G. 1984. Histoire de la France urbaine. Paris: le Seuil, 4 vols.
Dunbar, R.I.M. 1993. Co-evolution of cortex size, group size and language in humans. Behavioral
and Brain Sciences 16, 68194.
Durfee, E. 1988. Coordination of distributed problem solvers. Dordrecht: Kluwer.
Durfee, E.H., V R.Lesser, D.D.Corkill 1987. Coherent cooperation among distrib uted problem
solvers. IEEE Transactions on Computers 36, 1275291.
1989. Cooperative distributed problem solving. In The handbook of artificial intelligence, vol.
IV, A.Barr, P.R.Cohen, E.A.Feigenbaum (eds). Reading, Mass.: Addison-Wesley.
Durkheim, E. 1895. The rules of sociological method. Readings from Emile Durkheim. Chichester:
Ellis Horwood.
Egidi, M. 1992. Organizational learning and the division of labour. In Economics, bounded
rationality and the cognitive revolution, M.Egidi & R.Marris (eds). Aldershot: Edward Elgar.
Eibl-Eibesfeldt, I. 1967/1978. Grundri der vergleichenden Verhaltensforschung: Ethologie.
Munich: Piper.
Elster, J. 1982. Marxism, functionalism and game-theory: the case for methodological
individualism. Theory and Society 11, 45381.
Emerson, E.A. 1990. Temporal and modal logic. In Handbook of theoretical computer science,
J.van Leeuwen (ed.), 9961072. Amsterdam: Elsevier.
Emmeche, C. 1994. Is life a multiverse phenomena?. In Artificial life III, C.G.Langton (ed.), 553
68. Menlo Park, California: Addison Wesley.
Epstein, J.M. 1985. The calculus of conventional war: dynamic analysis without Lanchester theory.
Washington, DC: Brookings.
Ferber, J. 1989. Des objets aux agents. Doctoral thesis, University of Paris VI.
1991. Conception et programmation par objets. Paris: Herms.
Ferber, J. & P.Carle 1991. Actors and agents as reflective concurrent objects: a MERING IV
Perspective. IEEE Transactions on Systems, Man and Cybernetics.
Ferber, J. & A.Drogoul 1992. Using reactive multi-agent systems in simulation and problemsolving. In Distributed artificial intelligence: theory and praxis, N.M.Avouris & L. Gasser
(eds), 5380. ECSC-EEC-EAEC, Brussels and Luxembourg.
Fik, T.J. 1988. Hierarchical interaction: the modelling of a competing central place system. Annals
of Regional Science 22, 4869.
Fik, T.J. & G.F.Mulligan 1990. Spatial flows and competing central places: towards a general
theory of hierarchical interaction. Environment and Planning A 22, 52749.
Findler, N.V. 1987. Air traffic control: a challenge for artificial intelligence. AI Expert Magazine 2,
5966.
1989. Distributed knowledge-based systems in manufacturing. Keynote speech, Proceedings of the
international conference on artificial intelligence in engineering, Cambridge, England. Also
reprinted in Artiftcial intelligence in manufacturing, G.Rzevski (ed.), 1733. New York:
Springer.
1990. Contributions to a computer-based theory of strategies. New York: Springer.
1991. An artificial intelligence technique for fact retrieval: an application in medical
knowledge processing. Cambridge, Mass.: MIT Press.
1992. Distributed artificial intelligence: distributed planning and problem-solving, and multiagent systems. Invited tutorial, Proceedings of the first Singapore international conference on
intelligent systems, Singapore, 1992.
Findler, N.V. & J.Gao 1987. Dynamic hierarchical control for distributed problem-solving. Data
and Knowledge Engineering 2, 285301.
Findler, N.V. & Q.Ge 1989. Perceiving and planning before actingan approach to enhance global
network coherence. International Journal of Intelligent Systems 4, 45970.
Bibliography
244
Findler, N.V. & Q.Ge 1994. Distributed goal-oriented dynamic plan revision. International Journal
of Intelligent Systems 9, 183210.
Findler, N.V. & R.Lo 1986. An examination of distributed planning in the world of air traffic
control. Journal of Parallel and Distributed Processing 3, 41131.
1991. A distributed artificial intelligence approach to air traffic control IEE Proceedings, Part
D, Control theory and applications 138, 51524.
1993. Distributed air traffic controlPart I: Theoretical studies; and Part II: Explorations in a
testbed. Journal of Transportation Engineerig 119, 68194.
Findler, N.V. & U.K.Sengupta 1992. Multi-agent planning and collaboration in dynamic resource
allocation. Proceedings of the first international conference on AI planning systems, 3056,
University of Maryland, College Park.
Findler, N.V. & J.Stapp 1992. A distributed approach to optimized control of street traffic signals.
Journal of Transportation Engineering 118, 99110.
Finger, M., M.Fisher, R.Owens 1992. METATEM at work: modelling reactive systems using
executable temporal logic. Submitted to the Sixth international conference on industrial and
engineering applications of artificial intelligence and expert systems (IEA/AIE 93).
Switzerland: Gordon and Breach.
Fisher, M.1993. Concurrent METATEMa language for modelling reactive systems. In Parallel
architectures and languages, Europe (PARLE), Munich, Germany, 18596. Berlin: Springer.
Fisher, M. & H.Barringer 1991. Concurrent METATEM Processesa language for distributed AI.
In Proceedings of the European simulation multiconference, Copenhagen, Denmark.
Fisher, M. & R.Owens 1992. From the past to the future: executing temporal logic programs. In
Proceedings of logic programming and automated reasoning (LPAR), St Petersburg, Russia
(published in Lecture Notes in Computer Science 624).
Fisher, M. & M.Wooldridge 1993. Specifying and verifying distributed intelligent systems. In 6th
Portuguese conference on artificial intelligence (EPIA 93), 1328. Berlin: Springer.
Fletcher, R. 1986. Settlement in archaeology: world-wide comparison. World Archaeology 18, 59
83.
Floreano, D. 1993. Emergence of nest-based foraging strategies in ecosystems of neural networks.
In From animals to animats 2, J.A.Meyer, H.L.Roitblat, S.W.Wilson (eds). Cambridge, Mass.:
MIT Press.
Fodor, J. 1976. The language of thought. Sussex: Harvester Press.
Fontana, W. 1991. Functional self-organization in complex systems. In 1990 lectures in complex
systems, SFI studies in the sciences of complexity, vol. III, L.Nadel & D.Stein (eds), 4026.
Redwood City, California: Addison-Wesley
Forrest, S. 1990. Emergent computation. Physica D 42, 111.
Forrest, S. & J.H.Miller 1990. Emergent behaviour in classifier systems. In Emergent computation,
S.Forrest (ed.), Cambridge, Mass.: MIT Press.
Frankhauser, P. 1990. Beschreibung des Evolution urbaner Systeme mit derMastergleichung. Ph.D.
thesis, University of Stuttgart.
Fresneau, D., B.Corbara, J.P.Lachard 1989. Organisation sociale et structuration spatiale autour de
Couvain chez Pachycondyla apicalis (Formicidae, Ponerinae). Actes Collectif Insectes Sociaux
5, 8392.
Freyd, J. 1983. Shareability: the social psychology of epistemology. Cognitive Science 7, 191210.
Gamble, C. 1991. The social context for European Palaeolithic art. Proceedings of the Prehistoric
Society 57, 35.
Gasser, L., C.Braganza, N.Hermann 1987. MACE: a flexible testbed for distributed Al research. In
Distributed artificial intelligence., M.Huhns (ed), 11952. San Mateo, California: Morgan
Kaufmann.
Geertz, C. 1973. Thick description: toward an interpretive theory of culture. In The interpretation
of cultures, 330. New York: Basic Books.
Bibliography
245
Bibliography
246
1986. Escaping brittleness: the possibilities of general purpose learning algorithms applied to
parallel rule-based systems. In Machine learning II, R.S.Michalski, J. G.Carbonell,
T.M.Mitchell (eds). Los Altos, California: Morgan Kaufmann.
Holland, J.H., K.J.Holyoak, R.E.Nisbett, P.R.Thagard 1986. Induction: processes of inference,
learning and discovery. Cambridge, Mass.: MIT Press.
Huberman, B.A. & N.S.Glance 1993. Diversity and collective action. In Interdisciplinary
approaches to complex nonlinear phenomena, H.Haken & A.Mikhailov (eds). New York:
Springer.
Huff, J.O. 1976. A hierarchical migration model of population redistribution within a central place
hierarchy. Geographical Analysis 8, 23154.
Hutchins, E. 1980. Culture and inference. Cambridge, Mass.: Harvard University Press.
1990. The technology of team navigation. In Intellectual teamwork: social and technical bases
of cooperative work, J.Galegher, R.Kraut, C.Egido (eds), 191219. Hillsdale, New Jersey:
Lawrence Erlbaum Associates.
1991. Social organization of distributed cognition. In Perspectives on socially shared
cognition, L.Resnick, J.Levine, S.Teasley (eds), 283387. Washington, DC: The American
Psychological Association.
1993. How a cockpit remembers its speed. Unpublished manuscript, University of California,
San Diego.
in press. Cognition in the wild. Cambridge, Mass.: MIT Press.
Hutchins, E. & B.Hazlehurst 1991. Learning in the cultural process. In Artificial life II, C.Langton,
C.Taylor, D.Farmer, S.Rasmussen (eds), 689706. Redwood City, California: Addison-Wesley.
Hutchins, E. & T.Klausen in press. Distributed cognition in an airline cockpit. In Cog-nition and
communication at work, Y.Engestrm & D.Middleton (eds). New York: Cambridge University
Press.
Huth, P. 1988. Extended deterrence and the prevention of war. New Haven, Connecticut: Yale
University Press.
Jen, E. (ed.) 1990. 1989 lectures in complex systems. Santa Fe Institute Studies in the Sciences of
Complexity. Redwood City, California: Addison-Wesley.
Johnson, A.W. & T.Earle 1987. The evolution of human societies: from foraging group to agrarian
state. Stanford, California: Stanford University Press.
Johnson-Laird, P.N. 1983. Mental models. Cambridge, Mass.: Harvard University Press.
Kahneman, D. & A.Tversky 1984. Prospect theory: an analysis of decisions under risk.
Econometrica 47, 26391.
Kahneman, D., P.Slovic, A.Tversky 1982.Judgement under uncertainty: heuristics and biases. New
York: Cambridge University Press.
Kauffman, S. 1988. The evolution of economic webs. In The economy as an evolving complex
system, P.W.Anderson & K.J.Arrow (eds), 12545. Reading, Mass.: Addison-Wesley.
Kaufman, L. & P.J.Rousseeuw 1990. Finding groups in data: an introduction to cluster analysis.
New York: John Wiley.
Kennedy, P. 1987. The rise and fall of the great powers. New York: Random House.
Kinny, D., M.Ljungberg, A.Rao, E.Sonenberg, G.Tidhar, E.Werner 1992. Planned team activity.
Proceedings of MAAMAW 92 conference., July 2931, S.Martino al Cimino (Viterbo), Italy.
Koza, John R. 1992. Genetic programming. Cambridge, Mass.: MIT Press.
Kreps, D., P.Milgrom, J.Roberts, R.Wilson 1982. Rational cooperation in the repeated prisoners
dilemma. Journal of Economic Theory 27, 24552.
Kugler, R.L. 1990. NATOs future rle in Europetoward a more political alliance in a stable
11/2 bloc system. Technical report R-3923-FF, The Rand Corporation.
Lachaud, J-P. & D.Fresneau 1987. Social regulation in ponerine ants. In From individual to
collective behavior in social insects, J.M.Pasteels & J-L.Deneubourg (eds), Experientia
Supplementium 54, 197217.
Lakoff, G. 1987. Women, fire, and dangerous things. Chicago: University of Chicago Press.
Bibliography
247
Lanchester, F.W. 1916. Aircraft in warfare: the dawn of the first arm. London: Constable. Key
chapters reprinted as: Mathematics in warfare. The world of mathematics, J. Newman (ed.)
1956. New York: Simon and Schuster.
Langton, G.C. (ed.) 1994. Artificial life III. Menlo Park, California: Addison Wesley
Levinson, S.. in press. Interactional biases in human thinking. In The social origins of intelligence,
E.Goody (ed.). New York: Cambridge University Press.
Lewis, D. 1969. Convention. Cambridge, Mass.: Harvard University Press.
Lomborg, B. in press. Game-theory s multiple agents: the iterated prisoners dilemma. In
Decentralised AI4, C.Castelfranchi & E.Werner (eds). Berlin: Springer.
Lumsden, C.J. & E.O.Wilson 1985. The relation between biological and cultural evolution. Journal
of Social and Biological Structures 8, 34359.
March, J.G. & H.A.Simon 1958. Organizations. New York: John Wiley.
Marengo, L. 1992. Coordination and organizational learning in the firm. Journal of Evolutionary
Economics 2, 31326.
Maruichi, T., M.Ichikawa, M.Tokoro 1990. Modelling autonomous agents and their groups. In
Decentralized AIproceedings of the first European workshop on modelling autonomous
agents and multi-agent worlds (MAAMAW), Y.Demazeau &J.P.Muller (eds), 21534.
Amsterdam: Elsevier/North-Holland.
McCain, R.A. 1979. Reflections on the cultivation of taste. Journal of Cultural Economics 3, 30
52.
1981. Cultivation of taste, catastrophe theory, and the demand for works of art. American
Economic Review 71, 3324.
1982. Optimal subsidies to the arts in a shortsighted world. Journal of Cultural Economics 6,
1532.
1986. Game theory and the cultivation of taste. Journal of Cultural Economics 10, 115.
1991a. Groping: toward a behavioral metatheory of choice. In Handbook of behavioral
economics, R.Franz, H.Singh, J.Gerber (eds). Greenwich, Connecticut: JAI Press, 495508.
1991b. Deontology, consequentialism, and rationality. Review of Social Economy 49, 16895.
1991c. A linguistic theory of rationality. Social Science Information 30, 23355.
1992a. A framework for cognitive economics. New York: Praeger Publishers.
1992b. Cultivation of taste and bounded rationality: overview. Third ICARE conference on the
economics of the arts, Venice, Italy (December).
McCain, R.A. Cultivation of taste and bounded rationality: some computer simulations. Journal of
Cultural Economics, in press.
McClelland, J. 1988. Explorations in parallel distributed processing. Cambridge, Mass.: MIT
Press.
McClelland, J., D.Rumelhart, the PDP Group 1986. Parallel distributed processing: vol. 2,
explorations in the microstructure of cognition. Cambridge, Mass.: MIT Press.
Mellars, P.A. 1985. The ecological basis of social complexity in the Upper Paleolithic of
southwestern France. In Prehistoric hunter-gatherers: the emergence of cultural complexity, T.
D.Price & J.A.Brown (eds), 27197. New York: Academic Press.
Menczer, F. & D.Parisi 1992. Recombination and unsupervised learning: effects of crossover in
the genetic optimization of neural networks. Network 3, 42442.
Mercier, P. 1968. Anthropologie sociale et culturelle. Ethnologie Gnrale, 8871036. Paris: Ed.
Gallimard.
Minsky, M. 1985. Society of the mind. New York: Simon and Schuster.
Mokken, R.J. & F.N.Stokman 1985. Legislative analysis: methodology for the analysis of groups
and coalitions. In Coalition formation, H.A.M.Wilke (ed.). Amsterdam: Elsevier.
Moriconi, F.. 1993. Lurbanisation du monde, 19501990. Paris: Economica.
Morrill, R.L. 1962. Simulation of central place pattern over time. In Proceedings of the IGU
symposium in urban geography, K.Norborg (ed.), 109120. Lund: Glerup.
Bibliography
248
Bibliography
249
Schelling, T.C. 1960. The strategy of conflict. Oxford: Oxford University Press.
Schumpeter, J. 1934. Theory of economic development. Cambridge, Mass.: Harvard University
Press.
Shoham, Y. 1989. Time for action: on the relation between time, knowledge and action. In
Proceedings of the eleventh international joint conference on artificial intelligence (IJCAI),
Detroit, USA, 9549.
1990. Agent oriented programming. Technical Report STAN-CS-133590, Department of
Computer Science, Stanford University, California, USA.
Shoham, Y. & M.Tenneholtz 1992a. On the synthesis of useful social laws for artificial agent
societies (preliminary report). Proceedings of the AAAI Conference, 27681.
1992b. Emergent conventions in multi agent systems: Initial experimental results and
observations. Proceedings of the 3rd International Conference on KR&R. Cambridge, Mass.,
22532.
Shore, B. 1991. Twice-born, once conceived: meaning construction and cultural cognition.
American Anthropologist 93, 927.
Simon, H.A. 1962. The architecture of complexity. Proceedings of the American Philosophical
Society 106, 46782.
1981. The sciences of the artificial. Cambridge, Mass.: MIT Press.
Slobin, D. 1985. Crosslinguistic evidence for the language-making capacity. In The crosslinguistic
study of language acquisition, D.Slobin (ed.), 1157260. Hillsdale, New Jersey: Lawrence
Erlbaum.
Sobel, J. 1992. How to count to one thousand. Economic Journal 102, 18.
Sperber, D. 1985. Anthropology and psychology: towards an epidemiology of representations. Man
20, 7389.
1987. Les sciences cognitives, les sciences sociales et le matrialisme, in Introduction aux
sciences cognitives, 39446. Paris: Ed. Gallimard.
Stein, D.L. (ed.). 1989. Lectures in the sciences of complexity. Redwood City, California: AddisonWesley.
Stein, D. & L.Nadel (eds) 1991.1990 lectures in complex systems. Santa Fe Institute Studies in the
Sciences of Complexity. Redwood City, California: Addison-Wesley.
Stigler, G. and G.Becker 1977. De gustibus non est disputandum. American Economic Review 67,
6790.
Terna, P. 1993. Imitazione nelle reti neurali. Sistemi Intelligenti 5, 16973.
The raulaz, G., G.Goss, J.Gervet, J-L.Deneubourg 1990. Task differenciation in Polistes wasp
colonies: a model for self-organizing groups of robots. In Simulation of adaptive behaviour:
from animals to animats, J-A.Meyer & S.W.Wilson (eds), 34655. MIT Press/Bradford Books.
Thom, R. 1972. Stabilite structurelle et morphogenese: Essai dune theorie generale des modeles.
New York: Benjamin.
Thomas, B., Y.Shoham, A.Schwartz, S.Kraus 1991. Preliminary thoughts on an agent description
language. International Journal of Intelligent Systems 6, 497508.
Tilly, C. 1985. War making and state making as organized crime. In Bringing the state back in,
P.B.Evans, D.H.Rueschemeyer, T.A.Skocpol (eds). Cambridge: Cambridge University Press.
1990. Coercion, capital, and European states, AD 9901990. Oxford: Basil Blackwell.
Tobler, W.R. 1979. Cellular geography. In Philosophy of geography, S.Gale & G.Olsson (eds),
279386. Dordrecht: Reidel.
Trivers, R.L. 1985. Social evolution. Menlo Park, California: Benjamin Cummings.
Tuomela, R. 1991. We will do it: an analysis of group-intentions. Philosophy and
Phenomenological Research 51, 2497 7.
1992. Group beliefs. Synthese 91, 285318.
Vines, G. 1993. Planet of the free apes? New Scientist 138, 5 June 1993, 3942.
von Hayek, F.A. 1948. Economics and knowledge. In Individualism and economic order. Chicago:
University of Chicago Press.
Bibliography
250
Index
abstraction 123
action 147, 155, 215
adaptive cognition 10
adaptive learning 51
agency 8
aggression 253
alliance 21, 110, 113, 212
alliance, strategic 225
alternative futures 121
altruism 210, 239
ant 192
anthropomorphism 95
archaeology 2
artificial ethology 205
artificial intelligence 88
artificial life 2
artificial society 2, 6, 12
attachment behaviour 239
autoassociator 162
awareness 10
axiomatization 61
back-propagation of error 164
bias, sub-cognitive 10
bucket-brigade mechanism 56
cellular automata 88, 147
central place theory 87
chaos 123
chromosome 128
classifier systems 51
clobbering 106
co-evolution 41, 148
co-operation 6, 10, 210, 266, 268
co-ordination 7, 40, 49, 110
coalition formation 235
coercion 21
cognition 104, 121, 157
adaptive 10
cognitive agent 157, 266
cognitive anthropology 158
cognitive capabilities 49
cognitive capacity 137
Index
253
Index
exploratory simulation 4
254
Index
methodological holism 82, 145, 152
methodological individualism 5, 82, 146, 152
methodology 4, 121
micro-macro link 11, 88, 145
migration 95
model, social 105
multi-agent modelling 88, 190
multiple agent plan 105
mutation 53, 84, 128, 148
mutual aid 265
natural selection 84
neo-classical economics 42
neural net 2, 162, 240
norm 145, 212, 226, 252
norm generation 228
normative reasoning 266
object-orientation 192
optimization 3
permutation group 60
plan, multiple agent 105
planning 42, 105, 108
politics 2, 19
prehistory 104
prestige 105
Prisoners Dilemma 6, 22
problem-solving 43
production 94
public hidden units 187
random walk 127
rational choice 20
rationality 126
reactive system 269
reflexivity 144
representation 159, 183
mental 104, 266
returns to scale 45
rule of interaction 64
rule
condition-action 52
heuristic 126
kinship 60
utilitarian 259
selection 148
settlement system 86, 91
shared symbols 183
255
Index
simulation
exploratory 4
multi-agent 190
stochastic 25
Smalltalk 95
social behaviour 82, 239
social cognition 157
social complexity 103
social conflict 265
social dynamics 104
social group 246
social impact 148
social interaction 183
social learning 3
social model 105, 153
social representation 145
social strategy 3
social structure 147, 204, 238
sociality 4
sociogenesis 190, 195, 205
sociology 144, 238
sociotomies 209
space 239
spatial distribution 86
spatial relationships 59
stochastic simulation 25
stock market 213
strategic alliance 225
structural functionalism 5, 11
structuration 146, 152, 154
structure 147, 155
structuring properties 147
sub-cognitive bias 10
symbol 157
symbolic representation 183
system, settlement 91
target of simulation 2
teleological conservatism 131
teleologically conservative algorithm 132
temporal logic 269
theory building 122
theory
central place 87
conduit 181
economic 126
kin selection 239
structuration 154
tragedy of the commons 19
tribute model 21
256
Index
urbanization 86
utilitarian rule 259
utility 126, 252
utility function 133
validation 196
validity 121
257